Roadbed cavity intelligent detection method based on multi-feature fusion and deep learning and related equipment

Abstract

The invention relates to the technical field of roadbed cavity detection, in particular to a roadbed cavity intelligent detection method based on multi-feature fusion and deep learning and related equipment. The method comprises the following steps: acquiring road-base pulse signal data and preprocessing to form a preprocessed signal sequence; multi-dimensional features such as amplitude attenuation and frequency dispersion change are extracted based on a cavity physical effect model, and a high-discrimination fusion feature vector is generated through feature fusion and dimension reduction; and training a deep learning cavity recognition model by using the fusion features to realize automatic recognition of the cavity existence probability of the to-be-detected road section. And finally, spatial positioning and risk assessment are performed by combining cross-correlation analysis and frequency dispersion curve inversion, so that high-precision detection and reliable judgment of the roadbed cavity are realized. According to the invention, the accuracy, efficiency and intelligent degree of roadbed cavity detection are improved.

Description

Technical Field

[0001] This application relates to the field of roadbed cavity detection technology, and in particular to an intelligent roadbed cavity detection method and related equipment based on multi-feature fusion and deep learning. Background Technology

[0002] In related technologies, with the acceleration of urbanization and the continuous expansion of infrastructure construction such as roads and railways, the stability and safety of roadbed structures have become important factors affecting public transportation safety. The presence of cavities beneath the roadbed can lead to pavement settlement, cracking, and even collapse, seriously threatening driving safety and the lifespan of engineering projects. Therefore, early high-precision detection of cavities has become an important research direction in the field of infrastructure maintenance.

[0003] Traditional detection methods have certain limitations. Existing ground vibration detection relies on manual interpretation of spectral characteristics (such as the H / V ratio), which is easily influenced by subjective judgment and results in a high rate of missed detection for small-scale cavities. Current technology uses single-feature analysis, which, relying solely on frequency domain features, is insufficient to distinguish cavity signals from interference sources such as underground pipelines and rock strata interfaces. Urban roadbed vibration noise can mask weak cavity signals, and traditional filtering methods lead to the loss of effective signals.

[0004] Existing technologies have weak feature representation capabilities; a single H / V spectral ratio feature cannot fully characterize the nonlinear and non-stationary characteristics of signals caused by cavities, leading to missed detections or false alarms. Traditional methods rely on manual feature extraction and threshold judgment, which is difficult to meet the automation requirements of large-scale road inspections and has insufficient real-time processing capabilities. In addition, methods based on fixed parameters or rules are difficult to adapt to the cavity detection needs under different geological conditions, and the models have poor generalization ability.

[0005] In summary, the technical problems existing in the relevant technologies need to be improved. Summary of the Invention

[0006] The main objective of this application is to propose an intelligent detection method and related equipment for roadbed cavities based on multi-feature fusion and deep learning, aiming to improve the accuracy, efficiency and intelligence of roadbed cavity detection.

[0007] To achieve the above objectives, one aspect of this application proposes an intelligent detection method for roadbed voids based on multi-feature fusion and deep learning, the method comprising: Acquire ground vibration signal data for the target roadbed area; The ground pulsation signal data is subjected to frequency domain filtering and adaptive noise suppression to form a preprocessed signal sequence; Based on the physical effect model of holes, feature extraction is performed on the preprocessed signal sequence to generate a multi-dimensional feature set; Perform feature fusion and dimensionality reduction processing on the multi-dimensional feature set to obtain a fused feature vector; The deep learning hole identification model is optimized based on the fused feature vector; The preprocessed signal sequence of the road segment to be tested is input into the optimized deep learning hole identification model to obtain the probability of the existence of holes; Based on the probability of the cavity's existence, and combined with cross-correlation analysis and dispersion curve inversion results, the spatial location and risk assessment of the cavity are completed, resulting in cavity detection results, which include spatial location or high-risk early warning information.

[0008] In some embodiments, the physical effect model includes: The scattering effect sub-model is based on the Gaussian attenuation model of the scattering intensity in the preprocessed signal sequence, which is based on the change in distance between the station and the cavity center. The resonance effect sub-model identifies the energy accumulation characteristics of specific frequency components in the preprocessed signal sequence based on the difference in physical properties between the cavity and the surrounding medium. The wave velocity anomaly effect sub-model characterizes the low-speed shift of the phase velocity in the target frequency band in the preprocessed signal sequence based on the equivalent stiffness change caused by voids.

[0009] In some embodiments, the formula for the scattering effect sub-model is as follows: ; in, Indicates the first The intensity of scattering effect observed at each station; This represents the maximum effect amplitude at the center of the cavity; Indicates the first The location of each station; Indicates the location of the center of the cavity; The standard deviation represents the range of control effect attenuation.

[0010] In some embodiments, the multi-dimensional feature set includes: Temporal and nonlinear characteristics are used to quantify scattering effects, including root mean square, peak value, skewness, kurtosis, sample entropy, and Lyapunov exponent. Frequency domain characteristics are used to quantify the resonance effect, including the dominant frequency, bandwidth, spectral entropy, and H / V spectral ratio obtained through fast Fourier transform. Time-frequency domain features are used to quantify wave velocity anomaly effects, including energy concentration region features extracted through wavelet transform.

[0011] In some embodiments, the deep learning hole identification model includes a convolutional neural network model, a long short-term memory network model, and an attention model.

[0012] In some embodiments, optimizing the deep learning hole identification model based on the fused feature vector includes the following steps: Based on the fused feature vector, multi-scale convolution extraction is performed through the convolutional neural network model to form a local feature map; the local feature map characterizes scattering anomalies, resonant frequency band changes, and wave velocity disturbances. The local feature mapping is input into the long short-term memory network model for temporal correlation modeling to obtain a hidden state sequence that reflects the evolution law of the void physical effect; Based on the hidden state sequence, the attention model is used to perform feature weighting to obtain a global representation vector that highlights key physical responses; Based on the weighted feature representation, a feature representation layer and parameter framework for a deep learning hole recognition model are constructed. The parameters of each network structure in the deep learning hole recognition model are trained and converged using labeled samples.

[0013] In some embodiments, the process of spatially locating and assessing the risk of a cavity based on the probability of its existence, combined with cross-correlation analysis and dispersion curve inversion results, includes: The horizontal position of the cavity is determined based on the peak delay of the inter-station cross-correlation function, according to the probability of the cavity's existence. Based on the phase velocity anomaly obtained by inverting the dispersion curve, the characteristic period is converted into the burial depth of the cavity; By spatially combining the horizontal position with the burial depth, the three-dimensional positioning result of the cavity is obtained; A comprehensive index is constructed based on the intensity of scattering effect, the decrease in wave velocity, and the amplitude of dispersion anomaly. Based on the aforementioned comprehensive indicators, the affected area of ​​the cavity is delineated and risk is classified, and a structured detection report is output. The structured detection report includes the three-dimensional location of the cavity, its affected area, and its risk level.

[0014] In some embodiments, the formula for the cross-correlation function is as follows: ; in, This indicates that the time delay between station i and station j is... The cross-relationship number; i and j represent the location indices of the two stations, respectively; This represents the ground pulsation signal value of station i at time t; This represents the ground pulsation signal value of station j at time t+τ; Indicates time delay; N represents signal length; and This represents the standard deviation of the signals from stations i and j.

[0015] To achieve the above objectives, another aspect of this application proposes an intelligent roadbed cavity detection system, the system comprising: A broadband seismograph array is used to acquire ground vibration signals in the target roadbed area; The signal processing module is used to execute a roadbed cavity intelligent detection method based on multi-feature fusion and deep learning to generate cavity detection results; A visualization terminal is used to visualize and display the cavity detection results, which include spatial location or high-risk early warning information.

[0016] To achieve the above objectives, another aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0017] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0018] To achieve the above objectives, another aspect of this application provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0019] The embodiments of this application include at least the following beneficial effects: This application provides a method and related equipment for intelligent detection of roadbed cavities based on multi-feature fusion and deep learning. This scheme performs frequency domain filtering and adaptive noise suppression on ground pulsation signal data, and on this basis, introduces multi-dimensional feature extraction based on a cavity physical effect model. Then, through feature fusion and dimensionality reduction processing, a fused feature vector is constructed, which realizes quantitative characterization of cavity response characteristics and compression of redundant information. Compared with traditional methods that rely only on single time-domain or frequency-domain features, it improves the discriminative power of features in determining the existence of cavities and the robustness to complex interference in the field. Based on the fused feature vector, a deep learning cavity identification model is trained, and the identification result is output in the form of the probability of cavity existence. This avoids the problem of manually setting a single discrimination threshold based on experience. It can adaptively learn the cavity response patterns under different roadbed structures, working conditions, and survey line layouts, realizing high-precision, probabilistic identification of roadbed cavities, thereby reducing the risk of missed and false detections and improving the reliability and applicability of roadbed cavity detection. Based on the probability of cavity existence, cross-correlation analysis and dispersion curve inversion results are introduced to accurately locate suspected cavities in space. Risk assessment is then conducted in conjunction with probability information to construct an integrated diagnostic process of "probability identification - spatial inversion - risk classification". This process can comprehensively determine the location, size, and impact on the safety of the roadbed structure of cavities, providing a quantitative basis for subsequent reinforcement and maintenance decisions. This is conducive to improving the precision and proactive prevention capabilities of roadbed operation and maintenance management. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a method for intelligent detection of roadbed cavities based on multi-feature fusion and deep learning, provided in an embodiment of this application. Figure 2 This is a schematic diagram of the time-domain waveform of the original ground pulsation noise signal (first 5 seconds) provided in the embodiments of this application; Figure 3 This is a graph showing the analysis results of the cross-correlation function between different sampling stations provided in the embodiments of this application; Figure 4 This is an analysis diagram of the dispersion curve of ground vibration noise provided in the embodiments of this application; Figure 5 This is a schematic diagram of a roadbed cavity intelligent detection system provided in an embodiment of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0022] It is understood that the terms “first,” “second,” etc., used in this application may be used herein to describe various concepts, but unless otherwise stated, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of the embodiments of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the words “if,” “when,” or “in response to a determination” as used herein may be interpreted as “when…” or “when…” or “in response to a determination.”

[0023] As used in this application, the terms "at least one", "multiple", "each", "any", etc., "at least one" includes one, two or more, "multiple" includes two or more, "each" refers to each of the corresponding multiples, and "any" refers to any one of the multiples.

[0024] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0025] This application provides a method and related equipment for intelligent detection of roadbed cavities based on multi-feature fusion and deep learning. This scheme performs frequency domain filtering and adaptive noise suppression on ground pulsation signal data, and then introduces multi-dimensional feature extraction based on a cavity physical effect model. After feature fusion and dimensionality reduction, a fused feature vector is constructed, achieving quantitative characterization of cavity response characteristics and compression of redundant information. Compared with traditional methods relying solely on single time-domain or frequency-domain features, this improves the discriminative power of features in determining the presence or absence of cavities and its robustness to complex on-site interference. A deep learning cavity identification model is trained based on the fused feature vector, outputting the identification result in the form of a probability of cavity presence. This avoids the problem of manually setting a single discrimination threshold based on experience. It can adaptively learn cavity response patterns under different roadbed structures, working conditions, and survey line layouts, achieving high-precision, probabilistic identification of roadbed cavities, thereby reducing the risk of missed and false detections and improving the reliability and applicability of roadbed cavity detection. Based on the probability of cavity existence, cross-correlation analysis and dispersion curve inversion results are introduced to accurately locate suspected cavities in space. Risk assessment is then conducted in conjunction with probability information to construct an integrated diagnostic process of "probability identification - spatial inversion - risk classification". This process can comprehensively determine the location, size, and impact on the safety of the roadbed structure of cavities, providing a quantitative basis for subsequent reinforcement and maintenance decisions. This is conducive to improving the precision and proactive prevention capabilities of roadbed operation and maintenance management.

[0026] This application provides an intelligent detection method and related equipment for roadbed cavities based on multi-feature fusion and deep learning, relating to the field of roadbed cavity detection technology. The intelligent detection method for roadbed cavities based on multi-feature fusion and deep learning provided in this application can be applied to a terminal, a server, or software running on a terminal or server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, or vehicle terminal, but is not limited thereto; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the server can also be a node server in a blockchain network; the software can be an application implementing an intelligent detection method for roadbed cavities based on multi-feature fusion and deep learning, but is not limited to the above forms.

[0027] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0028] Please see Figures 1-5 As shown, this invention relates to a method and related equipment for intelligent detection of roadbed cavities based on multi-feature fusion and deep learning.

[0029] Figure 1 This is an optional flowchart of a roadbed cavity intelligent detection method based on multi-feature fusion and deep learning provided in an embodiment of this application. Figure 1 The method may include, but is not limited to, steps S1 to S7: S1: Acquire ground vibration signal data of the target roadbed area; In this embodiment, to ensure that the ground vibration signals can fully reflect the scattering, resonance, and wave velocity anomalies of the non-uniformity and potential cavities in the medium beneath the roadbed, five broadband seismometers are deployed along the road section to be measured. They are arranged in a single-line array near the roadbed centerline, with a fixed spacing of 20 m between each instrument. The total array length covers a 60 m cavity-sensitive area. The sampling rate of each seismometer is set to 100–200 Hz to ensure high-fidelity recording of the 0.1–10 Hz ground vibration characteristic frequency band. During sensor deployment, it is necessary to ensure tight coupling between the bottom of the sensor and the ground. In soft soil or gravel sections, the coupling interface must be cleaned beforehand, and light pressure or micro-piling methods are used to enhance the coupling quality, reducing the impact of environmental vibration and contact loosening on the recording quality.

[0030] Before formal data acquisition, all seismometers underwent consistency calibration. During calibration, five sensors were placed side-by-side on the same rigid base, and stable background vibration signals were recorded for at least two minutes. The cross-correlation coefficients between each sensor were calculated to confirm the consistency of their amplitude and phase responses. Formal data acquisition could only proceed if all cross-correlation coefficients were greater than 0.95; if they were below this threshold, recalibration or sensor replacement was necessary. The purpose of this calibration step was to ensure the reliability of the time delay characteristics in subsequent array cross-correlation analysis; otherwise, time delay errors would directly affect the inversion accuracy of the horizontal location of the cavity.

[0031] During formal data acquisition, each station synchronously records at least 10 minutes of continuous ground pulsation signals. The time-domain waveforms of each station are recorded synchronously (e.g.,...). Figure 2 As shown, Figure 2 The five sub-figures show the first five seconds of the signal. The acquired raw signal typically contains multiple spectral components: such as long-period background noise from ocean waves around 0.2 Hz, atmospheric excitation at 0.5 Hz, and Gaussian white noise disturbance. In addition, environmental noise sources such as road traffic and construction machinery need to be recorded simultaneously for noise template construction in the subsequent adaptive noise suppression model. Through the high-quality array acquisition described above, accurate and stable raw data input can be provided for subsequent frequency domain filtering, feature extraction, time-frequency analysis, and deep learning models.

[0032] It should be noted that geomagnetic pulsations are weak vibrations that continuously exist on the Earth's surface. Their signal sources come from natural phenomena such as ocean surges and atmospheric activity, and they are characterized by rich frequency bands (0.1-10Hz), uniform spatial distribution, and continuous free operation.

[0033] It should be noted that, in Figure 2 The image shows the first 5 seconds of the original ground pulsation signal recorded by five broadband seismometers deployed along the roadbed. Each sub-image corresponds to one station, and the signal contains Gaussian white noise, low-frequency components (0.2Hz), and mid-frequency components (0.5Hz). The signals from each station generally exhibit continuous, small-amplitude, and stable natural ground pulsation characteristics, with amplitudes limited to ±2, visually reflecting the background noise level of the acquisition environment and the effectiveness of the array's synchronous recording. Because the ground pulsation signal contains low- and mid-frequency components, its waveform density is high, its periodicity is weak, and it exhibits typical random vibration characteristics.

[0034] Based on the physical effect model of roadbed cavities, it can be seen that the scattering, resonance, and wave velocity anomalies introduced by the cavities to the surrounding medium modulate the signals of each station to varying intensities. The waveform distortion at station 3 in the figure is the most significant, manifested as local amplitude enhancement and waveform disturbance, which is a characteristic of enhanced scattering exhibited by stations near the cavity location. The waveforms of stations 1, 2, 4, and 5 are relatively stable, showing a regular difference in attenuation with distance. This figure provides the basic raw data for subsequent cross-correlation analysis, dispersion curve inversion, and multi-dimensional feature extraction, and is an important input for identifying cavity disturbance areas and constructing deep learning models.

[0035] S2: Perform frequency domain filtering and adaptive noise suppression on the ground pulsation signal data to form a preprocessed signal sequence; In this embodiment, the collected raw ground pulsation signals first enter the preprocessing stage.

[0036] (1) Frequency domain filtering: A bandpass filter is used to retain the low frequency band of 0.1–10 Hz, which contains the most significant hole resonance frequency and can significantly enhance the hole effect characteristics.

[0037] (2) Adaptive noise suppression: The environmental noise collected synchronously is modeled using an LMS adaptive filter to separate the noise components from the original signal, thereby achieving real-time suppression of traffic noise and construction machinery noise.

[0038] (3) Enhancement of the cavity sensitive frequency band: Based on the known cavity resonance center frequency (usually about 1–2 Hz), the frequency band is narrow-band filtered to make the spectral energy accumulation area caused by the cavity more prominent.

[0039] The preprocessed signal sequence waveforms are smoother and have a clearer frequency domain structure, providing high-quality input for subsequent physical feature extraction and deep learning.

[0040] S3: Based on the physical effect model of holes, feature extraction is performed on the preprocessed signal sequence to generate a multi-dimensional feature set; The physical effect models include: The scattering effect sub-model is based on Gaussian attenuation modeling of the scattering intensity in the preprocessed signal sequence according to the distance variation between the station and the cavity center. A resonance effect sub-model identifies the energy accumulation characteristics of specific frequency components in a preprocessed signal sequence based on the difference in physical properties between the cavity and the surrounding medium. The wave velocity anomaly effect sub-model is based on the equivalent stiffness change caused by voids to characterize the low-speed shift of phase velocity in the target frequency band in the preprocessed signal sequence.

[0041] The formula for the scattering effect sub-model is as follows: ; in, Indicates the first The intensity of scattering effect observed at each station; This represents the maximum effect amplitude at the center of the cavity; Indicates the first The location of each station; Indicates the location of the center of the cavity; The standard deviation represents the range of control effect attenuation.

[0042] The multidimensional feature set includes: Temporal and nonlinear characteristics are used to quantify scattering effects, including root mean square, peak value, skewness, kurtosis, sample entropy, and Lyapunov exponent. Frequency domain characteristics are used to quantify the resonance effect, including the dominant frequency, bandwidth, spectral entropy, and H / V spectral ratio obtained through fast Fourier transform. Time-frequency domain features are used to quantify wave velocity anomaly effects, including energy concentration region features extracted through wavelet transform.

[0043] In this embodiment, the physical effect model is constructed as follows: 1. Scattering effect sub-model (time domain and nonlinear characteristics): Physical Mechanism: Cavities, acting as strong scatterers, distort the seismic wave field, leading to more complex signal waveforms and inducing nonlinear dynamic behavior. Specifically, cavities reduce the support capacity of the roadbed, making it more susceptible to deformation under vibration; simultaneously, the nonlinear vibrations induced by cavities alter the signal distribution pattern. The intensity of this effect... Suitai Station Distance from the center of the cavity The increase in decreases with a Gaussian decay, as shown in the following formula: ; in, Indicates the first The intensity of scattering effect observed at each station; This represents the maximum effective amplitude at the center of the cavity (e.g., set to 0.4). Indicates the first The location of each station; Indicates the location of the cavity center (e.g., set to 60 meters); The standard deviation representing the range of control effect attenuation (e.g., set to 10 meters).

[0044] The features and extraction are as follows: Time-domain characteristics: Calculate the root mean square (RMS) and peak value of the signal to quantify the overall enhancement of vibrational energy and instantaneous pulse amplitude caused by scattering, which is manifested as an increase in RMS and peak value; calculate skewness and kurtosis to characterize the asymmetry and sharpness changes in waveform distribution caused by scattering distortion, i.e., anomalies in skewness (distribution asymmetry) and kurtosis (distribution sharpness).

[0045] Nonlinear characteristics: Sample entropy and Lyapunov exponent are calculated to characterize the increased signal complexity and altered chaotic properties of the system introduced by scattering effects. The physical mechanism is that the dynamic behavior of a roadbed system with voids is more complex and chaotic. The increase in sample entropy represents the increase in system complexity, while the Lyapunov exponent quantifies the system's sensitivity to initial conditions; that is, voids lead to a decrease in system stability.

[0046] 2. Resonance effect sub-model (frequency domain characteristics): Physical mechanism: As a low-velocity body, a void forms an impedance difference interface with the surrounding medium, capturing a specific frequency. The seismic wave energy forms a significant standing wave resonance.

[0047] The features and extraction are as follows: Frequency domain characteristics: The signal spectrum is obtained through Fast Fourier Transform (FFT), and the dominant frequency is extracted to identify the resonant frequency. The drift directly indicates changes in the resonant system; bandwidth is analyzed to assess resonance quality; spectral entropy is calculated to measure the degree of spectral energy concentration caused by resonance, and an increase in spectral entropy is a manifestation of the increased complexity of the system's resonant response. Simultaneously, the H / V spectral ratio (horizontal and vertical spectral ratio) is calculated, and a significant peak in its curve is a typical indicator of the void resonance effect.

[0048] 3. Sub-model of wave velocity anomaly effect (time-frequency domain characteristics): Physical mechanism: Voids reduce the stiffness of the medium, which in turn affects the phase velocity of seismic waves (especially surface waves). A significant drop occurred in a specific frequency band.

[0049] The features and extraction are as follows: Time-frequency domain characteristics: Wavelet transform (WT) is used to analyze the time-frequency distribution of the signal. The scattering and modulation effects of voids on the signal have time-varying characteristics. Resonance and wave velocity anomalies can both lead to the redistribution of signal energy at specific frequencies (corresponding to specific periods or depths). By extracting the characteristics of energy concentration areas, it is possible to effectively capture instantaneous energy bursts or continuous energy anomalies in specific frequency bands caused by voids, pinpoint the frequency bands where wave velocity anomalies occur, and provide key input for subsequent dispersion curve inversion, thereby locating low-velocity anomaly regions.

[0050] S4: Perform feature fusion and dimensionality reduction on the multi-dimensional feature set to obtain the fused feature vector; In this embodiment, a multi-dimensional feature set is obtained, including time-domain features, frequency-domain features, time-frequency-domain features, and nonlinear features, with dimensions typically exceeding 0-30. To avoid interference to the model caused by differences in the dimensions, scale, and noise sensitivity of different physical features, this embodiment first normalizes the multi-dimensional feature set and employs a variance stabilization strategy to suppress the influence of long-tail features on the overall distribution. Subsequently, feature correlation analysis is performed. Highly redundant features (e.g., RMS and peak-to-peak values ​​are highly correlated under scattering effects) are identified through the Pearson correlation coefficient matrix. Combined with key physical quantities such as cross-correlation distortion index and dispersion velocity reduction, a preliminary high-dimensional feature subset is constructed, making it more accurately reflect the essential modulation effects of scattering, resonance, and wave velocity anomalies in the void.

[0051] Based on this, Principal Component Analysis (PCA) is used to reduce the dimensionality of the selected multi-dimensional feature set. During the dimensionality reduction process, the feature covariance matrix is ​​calculated and its eigenvalues ​​are decomposed to obtain principal component vectors sorted by contribution rate. Principal components with a cumulative variance contribution rate exceeding 95% are retained to ensure that spatial information and physical constraints are not weakened. PCA can map the original high-dimensional features of over 20 dimensions to a low-dimensional fused feature space of 5–10 dimensions, while preserving core physical features such as scattering energy anomalies (e.g., RMS-sample entropy combination), resonant frequency band energy concentration (e.g., H / V spectral ratio-spectral entropy combination), and wave velocity anomaly periodic features (e.g., wavelet energy concentration scale), ensuring that the fused feature vector after dimensionality reduction still possesses void sensitivity.

[0052] The dimensionality-reduced fused feature vectors not only significantly reduce the computational burden on subsequent deep learning models but also improve the convergence speed and noise resistance of network training. Experiments in this embodiment demonstrate that the dimensionality-reduced fused feature vectors can further improve the stability and generalization ability of deep learning models in identifying holes, allowing the model to maintain robust performance in complex roadbed environments and under different geological conditions without over-reliance on a single resonant frequency band or a single scattering index. The resulting low-dimensional fused features serve as the standard input to the deep learning model, providing a more compact, high-information-density, and low-redundancy data foundation.

[0053] S5: Optimize the deep learning hole recognition model based on fused feature vectors; the deep learning hole recognition model includes a convolutional neural network model, a long short-term memory network model, and an attention model.

[0054] The optimization of the deep learning hole identification model based on fused feature vectors includes the following steps: S51: Based on the fused feature vector, multi-scale convolution extraction is performed through a convolutional neural network model to form a local feature map; the local feature map characterizes scattering anomalies, resonant frequency band changes, and wave velocity disturbances. S52: Input the local feature map into the long short-term memory network model to perform temporal correlation modeling and obtain the hidden state sequence that reflects the evolution law of the void physical effect; S53: Based on the hidden state sequence, an attention model is used to perform feature weighting to obtain a global representation vector that highlights key physical responses; S54: Feature representation layer and parameter framework for constructing a deep learning hole recognition model based on weighted feature representation; S55: The parameters of each network structure in the deep learning hole recognition model are trained and converged by using labeled samples.

[0055] In this embodiment, constructing a deep learning-based hole recognition model specifically includes building a hybrid model of a one-dimensional convolutional neural network (1D-CNN) and a long short-term memory network (LSTM). The 1D-CNN is used for local feature extraction, and the LSTM is used for temporal feature modeling. An attention model is introduced to enhance the weights of key features, and a transfer learning strategy is employed to accelerate the training process using a pre-trained model. The specific process and training strategy are as follows: (1) Model input and data preparation: The input is the fused feature vector. ,in The feature dimensions (e.g., including 20 features in total, such as time domain, frequency domain, and nonlinearity). The time series length is calculated by dividing 10 minutes of data into 5-second windows.

[0056] Data partitioning: The labeled dataset is randomly divided into training, validation, and test sets in a 7:2:1 ratio. The training set is used for model parameter learning, the validation set is used for hyperparameter tuning and early stopping, and the test set is used for final performance evaluation.

[0057] Data batch processing: Mini-batch gradient descent is used during training, with the batch size set to 32 to balance training efficiency and memory consumption.

[0058] (2) 1D-CNN local feature extraction: Structure: It contains three one-dimensional convolutional layers with 32, 64, and 128 kernels respectively, all with a kernel size of 3 and a stride of 1. Each convolutional layer is followed by a ReLU activation function and a max pooling layer with a kernel size of 2 and a stride of 2.

[0059] Forward propagation formula: Output feature map of the layer The calculation is as follows: ; in, Indicates the first Output feature maps after convolution and pooling; Indicates the first The weight parameters of the convolutional kernel; This represents a one-dimensional convolution operation; Indicates the first The output feature map of the layer is used as the input of this layer; Indicates the first The bias term parameters of the convolution kernel; The linear rectified activation function is defined as follows: ; This represents the max pooling operation, used for dimensionality reduction and preserving the translation invariance of features.

[0060] Function: Automatically learns and extracts the local correlations and abstract patterns of signals in the time dimension, such as the subtle distortions of short-time pulses or waveform envelopes caused by holes.

[0061] (3) LSTM temporal feature modeling: Structure: The feature sequence output by the CNN is input into a bidirectional LSTM layer with 128 hidden units.

[0062] Forward propagation formula: LSTM unit at each time step The calculation is as follows: ; in, Indicates the current time step; Indicates at time step The input vector; Indicates at time step The hidden state; Indicates at time step cellular state; The weight matrices corresponding to the forget gate, input gate, candidate cell state, and output gate, respectively, are obtained through training. The bias terms corresponding to the forget gate, input gate, candidate cell state, and output gate, respectively, are learned through training. This represents the Sigmoid activation function, which compresses the output to the (0,1) interval, controlling the degree to which the gate opens and closes; This represents the hyperbolic tangent activation function, which compresses the output to the (-1,1) interval to generate candidate cell states; This represents the Hadamard product, which is the element-wise multiplication of a matrix. This indicates the output of the forget gate, which controls the state of the previous cell. The proportion of information that needs to be forgotten; This represents the input gate output, controlling the state of candidate cells. The proportion of information that needs to be included; It represents the candidate cell state, containing new information that may be stored in the cell state at the current time step; This indicates the updated cell state, which combines the previous state with the current new information; This indicates the output of the output gate, controlling the current cell state. How much information needs to be output to the hidden state? It represents the hidden state of the current time step, and is also the output of that time step.

[0063] Bidirectional LSTM processes sequences in both forward and backward directions, and the final hidden state is the concatenation of the two.

[0064] Function: To capture the long-term dependencies and dynamic evolution patterns in ground pulsation signals, and to understand the temporal persistence of the cavity resonance effect.

[0065] (4) Feature weighting in attention model: Calculation process: Obtain the hidden state output of the LSTM at all time steps. Attention scores at each time step are calculated using a single-layer perceptron. : ; in, The weight matrix representing the attention model is obtained through training. This represents the bias term of the attention model, which is learned through training; The weights, representing the context vector of the attention model, are learned through training. The score is normalized to the attention weights using the Softmax function. : ; in, This represents the natural exponential function, used to map attention scores to positive values, ensuring that attention weights have probabilistic significance. represents the normalization factor, which is the sum of the exponential scores over all time steps; T represents the total number of time steps in the input sequence.

[0066] Final feature vector The (context vector) is a weighted sum of all hidden states: ; in, This indicates that the LSTM is at time step The hidden state.

[0067] Function: Enables the model to adaptively focus on key time segments most relevant to hole identification (e.g., the period of strongest resonance energy), thereby improving the model's discriminative ability and interpretability.

[0068] (5) Output layer and model training: Output layer: Context vector The input is fed into a fully connected layer, followed by a Softmax function, which outputs a two-dimensional probability vector. , representing the predicted probabilities of "no voids" and "voids", respectively.

[0069] Loss function: Cross-entropy loss is used to measure the difference between the predicted probability and the true label. Differences between (one-hot encoding): ; in, This represents the average loss value calculated for a batch of data. This indicates the batch size, which is the number of samples selected in one training session. Indicates the index of the sample in the batch; Indicates the first The true label of each sample is either 0 (no holes) or 1 (holes). The model predicts the first... The probability that a sample contains a hole.

[0070] Optimizer and Training: The Adam optimizer is used for parameter updates, with an initial learning rate set to 0.001. Early stopping is employed: training is terminated when the validation set loss no longer decreases for 10 consecutive epochs, and the parameters of the model that performed best on the validation set are restored to prevent overfitting.

[0071] (6) Data augmentation and training process: Synthetic data generation: Using a noise generation model, background ground pulsation signals with different signal-to-noise ratios and phase characteristics are synthesized in batches.

[0072] Data mixing: The synthesized noise signal is mixed with real signal samples labeled "no holes" in the dataset to generate new "no holes" training samples. At the same time, the synthesized noise is superimposed with the "with holes" signal at a certain ratio to simulate the behavior of the hole signal under different noise environments.

[0073] Training and Validation: The enhanced, expanded dataset is used for model training. This allows the model to learn more robust features, avoiding overfitting to the limited original data and significantly improving its accuracy and stability in detecting unknown road sections.

[0074] The calculation formula for the noise generation model is as follows: ; in, Indicates time Station Index The simulated background noise signal generated at the location; These represent the amplitude coefficients of Gaussian white noise, low-frequency, and mid-frequency components, respectively. Represents a standard Gaussian distributed random number; Characteristic frequencies representing background ground pulsations; Represents a time series; This indicates the station index, used to introduce the phase difference between stations.

[0075] (7) Transfer learning strategy: Pre-trained model: The model is pre-trained on a large public ground motion dataset or a cavity dataset with different geological conditions to obtain model parameters with basic signal representation capabilities.

[0076] Domain-Adaptive Fine-Tuning: For specific target road segments, the pre-trained model is fine-tuned using a small amount of labeled data. By incorporating domain-adaptive loss terms such as Maximum Mean Difference (MMD) into the loss function, the feature distribution difference between the source and target domains is minimized, enabling the model to quickly adapt to new scenarios and addressing the small sample size problem.

[0077] S6: Input the preprocessed signal sequence of the road section to be tested into the optimized deep learning hole identification model to obtain the probability of the existence of holes; S7: Based on the probability of the existence of a cavity, combined with the results of cross-correlation analysis and dispersion curve inversion, the spatial location and risk assessment of the cavity are completed, and the cavity detection results are obtained. The cavity detection results include spatial location or high-risk early warning information.

[0078] Specifically, based on the probability of cavity existence, combined with cross-correlation analysis and dispersion curve inversion results, the spatial location and risk assessment of the cavity are completed, yielding cavity detection results. These results include spatial location or high-risk early warning information, including: S71: Determine the horizontal position of the cavity based on the peak delay of the inter-station cross-correlation function according to the probability of the cavity's existence; S72: Based on the phase velocity anomaly obtained by inverting the dispersion curve, the characteristic period is converted into the burial depth of the cavity; S73: Spatially combine the horizontal position with the burial depth to obtain the three-dimensional location result of the cavity; S74: Construct a comprehensive index based on the intensity of scattering effect, the decrease in wave velocity, and the amplitude of dispersion anomaly; S75: Based on comprehensive indicators, the affected area of ​​the cavity is delineated and the risk is classified, and a structured detection report is output; the structured detection report includes the three-dimensional location of the cavity, the affected area and the risk level.

[0079] In this embodiment, based on the probability of a cavity's existence, when a cavity is determined to exist, the following comprehensive localization and risk assessment process is executed: (1) Spatial positioning: Based on the peak delay time of cross-correlation of the station array, the horizontal center position of the cavity is calculated using the direction of arrival or tomographic imaging method.

[0080] Specifically, the cross-correlation function of all station pairs is calculated, with the maximum lag time set to 2 seconds, and the cross-correlation matrix is ​​generated (e.g., Figure 3 As shown, Figure 3 (See sub-diagrams for Zhongtai Stations 1-2 to 4-5).

[0081] Figure 3 The paper presents the normalized cross-correlation function results of each pair of stations from the five seismographs, used to characterize the similarity and propagation delay characteristics of the ground pulsation signal between different stations. Overall, the cross-correlation curves of most station pairs (such as stations 1-2, 1-3, 1-4, 1-5, 2-4, 3-5, etc.) show low-amplitude, random fluctuations without obvious periodic structure. This is consistent with the broadband and random characteristics of the ground pulsation signal itself, indicating that these station pairs are not significantly affected by scattering disturbances or abnormal propagation paths, and the signal similarity is normal with stable propagation paths.

[0082] However, the key station pairs involving station 3 (especially the cross-correlation curves of stations 2-3, 3-4, and 2-5) exhibited significant periodic oscillation structures and peak enhancements, deviating markedly from the random cross-correlation characteristics of other station pairs. This was manifested as symmetry disruption, recurring peaks, and strong correlations concentrated around specific lag times. The cavity located near station 3 altered the propagation path of geostationary waves, causing observable anomalous delays and enhanced coherence in the cross-correlation functions of station pairs near the cavity. This figure clearly verifies the impact of the cavity on wavefield propagation, providing crucial physical evidence for subsequent cavity localization, wave velocity anomaly inversion, and deep learning model training.

[0083] The formula for the cross-correlation function is as follows: ; in, This indicates that the time delay between station i and station j is... The cross-relationship number; i and j represent the location indices of the two stations, respectively; This represents the ground pulsation signal value of station i at time t; This represents the ground pulsation signal value of station j at time t+τ; Indicates time delay; N represents signal length; and This represents the standard deviation of the signals from stations i and j.

[0084] Verification and Quantification: The peak time difference of the cross-correlation function was extracted, and the apparent velocity was calculated in conjunction with the station spacing to identify the anomalous wave propagation delay caused by scattering effects. In this embodiment, the pre-defined hole was located 60 meters from the first station, and its scattering effect followed a Gaussian attenuation model. The analysis results conclusively showed that the cross-correlation curve of the third station, which is closest to the hole center, exhibited the most significant distortion and anomalous delay, which perfectly matched the model prediction that "the most significant anomaly is observed at the nearest station." The degree of symmetry violation of the cross-correlation function was used as a key indicator for quantifying the strength of the scattering effect.

[0085] By combining the period corresponding to the wave velocity anomaly obtained by dispersion curve inversion, the burial depth of the cavity is estimated through half-wavelength depth conversion relationship or joint inversion algorithm.

[0086] Specifically, a theoretical dispersion curve is generated based on the background velocity model, and then compared with the dispersion curve obtained by inverting measured data.

[0087] Verification and Quantification: Through comparison, the low-velocity anomaly zone within a specific period caused by the wave velocity anomaly effect is accurately located (corresponding to...). Figure 4 (The area marked in red in the middle). Within the depth period corresponding to the preset cavity location (60 meters), the measured curve showed a significant low-velocity anomaly with a velocity decrease of more than 15%. This directly confirmed the physical mechanism of the cavity leading to a decrease in medium stiffness and completed the spatial location and quantitative assessment of the wave velocity anomaly effect.

[0088] Figure 4 The diagram shows the dispersion curves obtained from ground vibration noise inversion. The solid blue line represents the measured phase velocity variation with the period, while the dashed black line represents the background velocity model. Overall, the background velocity shows a smooth upward trend throughout the period, reflecting the normal roadbed medium's wave velocity gradually increasing with the period. However, the measured curve shows a significant dip in the period range of approximately 1.2–1.8 s, with the lowest phase velocity being about 15%–20% lower than the background velocity, forming a significant "low-velocity anomaly zone." This characteristic is consistent with the physical mechanism of cavities leading to a decrease in medium stiffness and a reduction in shear wave velocity, and is a typical dispersion manifestation of the cavity wave velocity anomaly effect.

[0089] The red shaded area in the figure marks the dominant frequency band of the cavity response, with a central period of approximately T = 1.5 s. The corresponding surface wave sensitive depth matches the cavity depth. The velocity anomaly within this period indicates that the surface pulsation signal is affected by a low-velocity weakening layer when propagating to this depth, significantly reducing its propagation velocity and thus forming an observable concave structure on the dispersion curve. This figure effectively verifies the modulation effect of the cavity on the surface wave propagation path and can serve as an important basis for the existence and depth location of the cavity.

[0090] The formula for calculating the dispersion curve is as follows: ; ; ; in, Represents the background phase velocity model; Indicates the observed phase velocity; The period is represented by 0.1-10s; 800m / s represents the bedrock velocity; 400m / s represents the velocity variation range; 1.5s represents the cavity response center period; 0.3s represents the cavity response bandwidth; and 0.2 represents the maximum velocity reduction percentage (20%).

[0091] By combining horizontal position and depth information, three-dimensional spatial positioning of the cavity can be achieved.

[0092] (2) Scope delineation and risk assessment: Using a scattering effect sub-model, with the calculated cavity center as the origin, the influence range of the cavity is delineated based on the spatial interpolation results of the signal distortion degree and wave velocity reduction amplitude.

[0093] Areas with a wave velocity reduction greater than 15% are automatically marked as "high-risk areas," and areas with a wave velocity reduction between 5% and 15% are marked as "medium-risk areas," and are highlighted in different colors on the visualization terminal.

[0094] (3) Recommendations for output and disposal of results: Generate a structured detection report that includes the three-dimensional coordinates of the cavity, its impact range, and its risk level.

[0095] Automatically outputs drilling verification suggestions, including suggested verification point coordinates (prioritizing the center of high-risk areas) and suggested drilling depth (based on the estimated depth with an additional 2-5 meters of safety margin).

[0096] Based on the positioning results, the system can recommend an optimized station deployment plan for the next step of encrypted measurement.

[0097] Please see Figure 5 This application also provides an intelligent detection system for roadbed cavities, the system comprising: A broadband seismograph array is used to acquire ground vibration signals in the target roadbed area; The signal processing module is used to execute a roadbed cavity intelligent detection method based on multi-feature fusion and deep learning to generate cavity detection results; A visualization terminal is used to visualize and display the cavity detection results, which include spatial location or high-risk early warning information.

[0098] It is understood that the content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0099] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0100] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0101] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0102] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0103] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0104] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented by the embodiments of this program product are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0105] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0106] This application provides a method and related equipment for intelligent detection of roadbed cavities based on multi-feature fusion and deep learning. This scheme performs frequency domain filtering and adaptive noise suppression on ground pulsation signal data, and then introduces multi-dimensional feature extraction based on a cavity physical effect model. After feature fusion and dimensionality reduction, a fused feature vector is constructed, achieving quantitative characterization of cavity response characteristics and compression of redundant information. Compared with traditional methods relying solely on single time-domain or frequency-domain features, this improves the discriminative power of features in determining the presence or absence of cavities and its robustness to complex on-site interference. A deep learning cavity identification model is trained based on the fused feature vector, outputting the identification result in the form of a probability of cavity presence. This avoids the problem of manually setting a single discrimination threshold based on experience. It can adaptively learn cavity response patterns under different roadbed structures, working conditions, and survey line layouts, achieving high-precision, probabilistic identification of roadbed cavities, thereby reducing the risk of missed and false detections and improving the reliability and applicability of roadbed cavity detection. Based on the probability of cavity existence, cross-correlation analysis and dispersion curve inversion results are introduced to accurately locate suspected cavities in space. Risk assessment is then conducted in conjunction with probability information to construct an integrated diagnostic process of "probability identification - spatial inversion - risk classification". This process can comprehensively determine the location, size, and impact on the safety of the roadbed structure of cavities, providing a quantitative basis for subsequent reinforcement and maintenance decisions. This is conducive to improving the precision and proactive prevention capabilities of roadbed operation and maintenance management.

[0107] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0108] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0109] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0110] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0111] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0112] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A method for intelligent detection of roadbed voids based on multi-feature fusion and deep learning, characterized in that, The method includes: Acquire ground vibration signal data for the target roadbed area; The ground pulsation signal data is subjected to frequency domain filtering and adaptive noise suppression to form a preprocessed signal sequence; Based on the physical effect model of holes, feature extraction is performed on the preprocessed signal sequence to generate a multi-dimensional feature set; Perform feature fusion and dimensionality reduction processing on the multi-dimensional feature set to obtain a fused feature vector; The deep learning hole identification model is optimized based on the fused feature vector; The preprocessed signal sequence of the road segment to be tested is input into the optimized deep learning hole identification model to obtain the probability of the existence of holes; Based on the probability of the cavity's existence, and combined with cross-correlation analysis and dispersion curve inversion results, the spatial location and risk assessment of the cavity are completed, resulting in cavity detection results, which include spatial location or high-risk early warning information.

2. The method according to claim 1, characterized in that, The physical effect model includes: The scattering effect sub-model is based on the Gaussian attenuation model of the scattering intensity in the preprocessed signal sequence, which is based on the change in distance between the station and the cavity center. The resonance effect sub-model identifies the energy accumulation characteristics of specific frequency components in the preprocessed signal sequence based on the difference in physical properties between the cavity and the surrounding medium. The wave velocity anomaly effect sub-model characterizes the low-speed shift of the phase velocity in the target frequency band in the preprocessed signal sequence based on the equivalent stiffness change caused by voids.

3. The method according to claim 2, characterized in that, The formula for the scattering effect sub-model is as follows: ； in, Indicates the first The intensity of scattering effect observed at each station; This represents the maximum effect amplitude at the center of the cavity; Indicates the first The location of each station; Indicates the location of the center of the cavity; The standard deviation represents the range of control effect attenuation.

4. The method according to claim 1, characterized in that, The multi-dimensional feature set includes: Temporal and nonlinear characteristics are used to quantify scattering effects, including root mean square, peak value, skewness, kurtosis, sample entropy, and Lyapunov exponent. Frequency domain characteristics are used to quantify the resonance effect, including the dominant frequency, bandwidth, spectral entropy, and H / V spectral ratio obtained through fast Fourier transform. Time-frequency domain features are used to quantify wave velocity anomaly effects, including energy concentration region features extracted through wavelet transform.

5. The method according to claim 1, characterized in that, The deep learning hole identification model includes a convolutional neural network model, a long short-term memory network model, and an attention model.

6. The method according to claim 5, characterized in that, The optimization of the deep learning hole identification model based on the fused feature vector includes the following steps: Based on the fused feature vector, multi-scale convolution extraction is performed through the convolutional neural network model to form a local feature map; the local feature map characterizes scattering anomalies, resonant frequency band changes, and wave velocity disturbances. The local feature mapping is input into the long short-term memory network model for temporal correlation modeling to obtain a hidden state sequence that reflects the evolution law of the void physical effect; Based on the hidden state sequence, the attention model is used to perform feature weighting to obtain a global representation vector that highlights key physical responses; Based on the weighted feature representation, a feature representation layer and parameter framework for a deep learning hole recognition model are constructed. The parameters of each network structure in the deep learning hole recognition model are trained and converged using labeled samples.

7. The method according to claim 1, characterized in that, Based on the probability of the cavity's existence, combined with cross-correlation analysis and dispersion curve inversion results, the spatial location and risk assessment of the cavity are completed, including: The horizontal position of the cavity is determined based on the peak delay of the inter-station cross-correlation function, according to the probability of the cavity's existence. Based on the phase velocity anomaly obtained by inverting the dispersion curve, the characteristic period is converted into the burial depth of the cavity; By spatially combining the horizontal position with the burial depth, the three-dimensional positioning result of the cavity is obtained; A comprehensive index is constructed based on the intensity of scattering effect, the decrease in wave velocity, and the amplitude of dispersion anomaly. Based on the aforementioned comprehensive indicators, the affected area of ​​the cavity is delineated and risk is classified, and a structured detection report is output. The structured detection report includes the three-dimensional location of the cavity, its affected area, and its risk level.

8. The method according to claim 7, characterized in that, The formula for the cross-correlation function is as follows: ； in, This indicates that the time delay between station i and station j is... The cross-relationship number; i and j represent the location indices of the two stations, respectively; This represents the ground pulsation signal value of station i at time t; This represents the ground pulsation signal value of station j at time t+τ; Indicates time delay; N represents signal length; and This represents the standard deviation of the signals from stations i and j.

9. A roadbed cavity intelligent detection system, characterized in that, The system includes: A broadband seismograph array is used to acquire ground vibration signals in the target roadbed area; A signal processing module is configured to perform the method described in any one of claims 1-8 to generate cavity detection results; A visualization terminal is used to visualize and display the cavity detection results, which include spatial location or high-risk early warning information.

10. An electronic device, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the method as described in any one of claims 1 to 8.

Images