Underground cable detection method and system based on real simulation and time sequence fusion

Abstract

The invention provides an underground cable detection method and system based on real simulation and time sequence fusion, relates to the field of ground penetrating radar and deep learning, and aims to completely capture continuous evolution of cable hyperbolic curve features along with depth through depth sequence modeling and enhance the weight of a strong reflection peak value through an attention mechanism. The problem of feature ambiguity caused by signal attenuation of the deeply-buried cable is solved, depth features of transverse adjacent positions are correlated through space-depth fusion, the reflection intensity of a single pixel can be observed through the arrangement, the change trend of a reflection signal along with the depth can be more concerned, and the measurement accuracy is improved. And meanwhile, physical and time sequence supervision signals are constructed through three-level labeling, specific electromagnetic characteristics of the cable are separated by using an attention mechanism, and an evolution law of a deep fusion module is combined, so that the accuracy and robustness of cable identification in a complex geological environment are improved. The core problems of strong clutter interference, missing detection of weak signals and missing of physical attributes in the prior art are solved.

Description

Technical Field

[0001] This invention relates to the fields of ground-penetrating radar and deep learning, and in particular to an underground cable detection method and system based on real-world simulation and time-series fusion. Background Technology

[0002] In the fields of ground-penetrating radar (GPR) detection and deep learning, accurately identifying and distinguishing underground live cables relies heavily on the feature capture capabilities of deep learning models and the scene coverage of training datasets. However, existing technologies face three major bottlenecks: First, GPR equipment is expensive, and traditional applications are mainly geological surveys, resulting in a scarcity of specialized image datasets for "live cables + complex underground interference." This lack of sufficient targeted training samples hinders the model's ability to adequately learn the features of live cables. Second, traditional GPR datasets mostly simulate single scenes, typically containing only unaffected pure soil and a single cable, failing to reproduce the real underground environment of cities—making it difficult to recreate interwoven plastic and concrete pipes. The simulation data, including pipeline networks, cannot simulate complex interference elements such as irregular stones and uneven soil stratification. This results in significant differences between the simulated data and the signals collected by actual GPR (Gas Perceptual Reflection) data. The model trained on the simulation data has weak generalization ability and is difficult to adapt to real detection scenarios. At the same time, the traditional YOLO model, as a general target detection architecture, has not been optimized for the characteristics of GPR images. Underground cables often appear as small targets in GPR images, and the reflected signals of deep cables exhibit "weak signal characteristics" due to soil attenuation, making them easily submerged by environmental noise. The reflected signals of interference such as stones and metal fragments are similar to the strong reflection characteristics of the cable surface, making it difficult for the model to effectively distinguish them. Ultimately, this leads to a high rate of missed detection for deep cables and a high rate of false detection for interfering targets. Summary of the Invention

[0003] To address the aforementioned technical issues, this invention proposes an underground cable detection method and system based on realistic simulation and temporal fusion. By constructing an underground model that more closely resembles the real-world scenario, the system performs temporal fusion on the scanned data to extract deeply correlated features. Simultaneously, physical constraints are added to the scanned data to effectively suppress the interference of noise in the detection data on the detection results, thereby effectively reducing the missed detection rate of layered cables and the false detection rate of interference targets.

[0004] The first aspect of this invention provides a method for detecting underground cables based on real-world simulation and temporal fusion. This method includes: forming an underground cable distribution model set using simulation software; performing ultrasonic B-scans on the underground cable distribution model set using the simulation software to obtain a scan dataset, the scan dataset including scanned images; performing deep fusion on the scan dataset using a trained data processing model to obtain a temporal fusion feature set; training a detection model using the temporal fusion feature set until training is complete; inputting actual detection scan data into the data processing model to obtain temporal fusion features; and inputting the temporal fusion features into the detection model to obtain a detection result.

[0005] In some implementations, the step of using a trained data processing model to perform deep fusion of the scanned dataset to obtain a temporal fusion feature set includes: processing the data in the scanned dataset through physical constraints to obtain a filtered scanned dataset; converting each training data in the filtered scanned dataset into a pixel sequence in the depth direction; processing the pixel sequence in the depth direction through a unidirectional long short-term memory network to obtain multiple hidden layer parameters; and weighting and fusing the hidden layer parameters through an attention mechanism to obtain the temporal fusion feature set.

[0006] In some implementations, the step of weightedly fusing the hidden layer parameters through an attention mechanism to obtain the temporal fusion feature set includes: calculating the importance score of each hidden layer parameter through a multilayer perceptron mechanism, normalizing the importance score to obtain the importance weight, and using the importance weight as the weight of the corresponding hidden layer parameter to perform weighted fusion of the hidden layer parameters.

[0007] In some implementations, processing the data in the scanned dataset through physical constraints to obtain a filtered scanned dataset includes: performing image recognition on the scanned images in the scanned dataset to obtain high-frequency data and low-frequency data; performing convolution calculation with the high-frequency data using an edge detection operator to obtain a high-frequency feature map; processing the low-frequency data using a dilated convolutional layer to obtain a low-frequency feature map; and performing weighted fusion of the high-frequency feature map and the low-frequency feature map using an attention mechanism to obtain the filtered scanned dataset.

[0008] In some implementations, the step of weightedly fusing the high-frequency feature map and the low-frequency feature map using an attention mechanism to obtain the screening scan dataset includes: performing max pooling on the high-frequency feature map to extract the maximum reflection intensity, reflecting the saliency of the metal core; performing average pooling on the low-frequency feature map to calculate the average texture energy, reflecting the saliency of the insulating layer; concatenating the maximum reflection intensity and the average texture energy through channels, compressing the concatenated result into one dimension through a one-dimensional convolutional layer, and obtaining weights through normalization; and weightedly fusing the high-frequency feature map and the low-frequency feature map using the weights to obtain the screening scan dataset.

[0009] In some implementations, the process of forming an underground cable distribution model set using simulation software includes: constructing a virtual model of the underground cable, setting cable models within the underground cable model; setting non-cable conduit models, irregular rock models, and air cavity models within the underground cable model; and injecting power frequency signals into the cable models through excitation.

[0010] A second aspect of this invention provides an underground cable detection system based on real-world simulation and temporal fusion. The underground cable detection system includes: a simulation module, configured to generate an underground cable distribution model set using simulation software, and further configured to perform ultrasonic B-scans on the underground cable distribution model set using the simulation software to obtain a scan dataset, the scan dataset including scanned images; a fusion module, configured to perform deep fusion of the scan dataset using a trained data processing model to obtain a temporal fusion feature set; a training module, configured to train a detection model using the temporal fusion feature set until training is complete and a detection model is obtained; and a detection module, configured to input actual detection scan data into the data processing model to obtain temporal fusion features, and input the temporal fusion features into the detection model to obtain detection results.

[0011] In some implementations, the fusion module is further configured to process the data in the scan dataset through physical constraints to obtain a filtered scan dataset; to convert each training data in the filtered scan dataset into a pixel sequence in the depth direction; to process the pixel sequence in the depth direction through a unidirectional long short-term memory network to obtain multiple hidden layer parameters; and to perform weighted fusion of each hidden layer parameter through an attention mechanism to obtain the temporal fusion feature set.

[0012] In some implementations, the fusion module is further configured to calculate the importance score of each hidden layer parameter through a multilayer perceptron mechanism, normalize the importance score to obtain the importance weight, and use the importance weight as the weight of the corresponding hidden layer parameter to perform weighted fusion of each hidden layer parameter.

[0013] In some implementations, the fusion module is further configured to perform image recognition on the scanned images in the scanned dataset to obtain high-frequency data and low-frequency data; to perform convolution calculation with the high-frequency data using an edge detection operator to obtain a high-frequency feature map, and to process the low-frequency data using a dilated convolutional layer to obtain a low-frequency feature map; and to perform weighted fusion of the high-frequency feature map and the low-frequency feature map using an attention mechanism to obtain the filtered scanned dataset.

[0014] The deep fusion module's adaptability to B-scan images is reflected in three aspects: First, through depth sequence modeling, it fully captures the continuous evolution of the cable's hyperbolic features with depth; second, the attention mechanism strengthens the weight of strong reflection peaks, solving the feature ambiguity problem caused by signal attenuation in deeply buried cables; and third, spatial-depth fusion establishes a correlation between depth features at laterally adjacent locations, aiding in the identification of complete hyperbolic morphology. The GPM-YOLO model designed above not only observes the reflection intensity of individual pixels but also focuses on the changing trend of the reflection signal with depth, thus significantly improving the accuracy and robustness of cable identification in complex geological environments. By constructing physical and temporal supervision signals through three-level annotation, using the attention mechanism to separate the cable's unique electromagnetic features, and combining the evolutionary laws of the deep fusion module, it solves the core problems of traditional YOLO models in GPR images: "strong clutter interference, missed weak signals, and missing physical attributes." Attached Figure Description

[0015] Figure 1 A flowchart illustrating an underground cable detection method based on real-world simulation time-series fusion, provided as an embodiment of the present invention; Figure 2 for Figure 1 A flowchart illustrating step S101; Figure 3 A diagram of the underground cable distribution model constructed for simulation modeling software; Figure 4 for Figure 1 A flowchart illustrating step S103; Figure 5 This is a flowchart illustrating the training process of the detection model provided in an embodiment of the present invention. Figure 6 This is a schematic diagram of the data processing flow of ground-penetrating radar data using GPM-YOLO, provided in an embodiment of the present invention. Figure 7 This is a schematic diagram of an underground cable detection system based on real-world simulation time-series fusion, provided as an embodiment of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0017] The specific technical features described in the various embodiments in the detailed implementation can be combined in various ways without contradiction. For example, different implementation methods can be formed by combining different specific technical features. In order to avoid unnecessary repetition, the various possible combinations of the specific technical features in this invention will not be described separately.

[0018] It should also be noted that, in order to avoid obscuring the present invention with unnecessary details, only the structures and / or processing steps closely related to the present invention are shown in the accompanying drawings, while other details that are not closely related to the present invention are omitted.

[0019] Additionally, it should be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. In the following description, the terms "first," "second," etc., are used merely to distinguish different objects and do not indicate any similarity or connection between them. It should be understood that the directional descriptions such as "above," "below," "inside," and "outside" refer to the orientation under normal use conditions.

[0020] In the following specific implementation, the underground cable detection method based on realistic simulation and temporal fusion constructs a training dataset in batches using simulation software. The detection model is then trained using this training dataset. Furthermore, a pre-trained data processing model is embedded within the training model. This model incorporates physical constraints and performs temporal fusion on the data from the training dataset. After training, the underground cable can be detected using both the trained data processing model and the trained detection model. In other words, the detection of underground cables can be achieved by inputting data acquired from actual radar at the ground level into these two models. The following examples illustrate this underground cable detection method.

[0021] In some embodiments, such as Figure 1 As shown, Figure 1 This is a flowchart illustrating an underground cable detection method based on real-world simulation and time-series fusion, provided by an embodiment of the present invention. The process of this underground cable detection method includes: Step S101: Use simulation software to create a set of underground cable distribution models.

[0022] This can be understood as using simulation software to construct virtual models of different underground scenarios and cable distributions. Different underground interference models can be set in these virtual models, such as underground non-cable pipelines, irregular rock models, or air cavities. By changing the distribution of underground cable models and underground interference models, a rich and diverse set of virtual models can be constructed.

[0023] Optional, such as Figure 2 As shown, the specific steps of step S101 include: Step S201: Construct a virtual model of the underground cable and set up the cable model in the underground cable model.

[0024] This can be understood as arranging the distribution locations of virtual cable models within an underground virtual model.

[0025] Step S202: Set up a non-cable conduit model, an irregular rock model, and an air cavity model in the underground cable model.

[0026] That is, underground interference targets are set in the virtual model of underground cables to simulate the complex underground environment.

[0027] Step S203: Inject power frequency signals into the cable model through excitation.

[0028] In addition to their shape, cables possess unique electromagnetic characteristics due to their function of transmitting electrical energy. These characteristics require additional excitation to imbue the cable model with specific electromagnetic features. For example, the process of injecting a power frequency signal includes: defining the energized excitation waveform using commands to generate a 50Hz, 220V sine wave to simulate the power frequency signal of an actual energized cable. At this point, the cable insulation layer will exhibit periodic oscillating textures due to electromagnetic induction, becoming a key feature for subsequent model identification of "energized cables"; loading the excitation signal onto the cable core, with the sine wave transmitted through a 50Ω characteristic impedance transmission line to the corresponding coordinate position of the cable core in the underground model, ensuring uniform signal transmission along the cable's extension direction; and marking grounding boundaries at both ends of the cable's metal sheath using boundary definition commands to prevent the excitation signal from leaking into the soil and causing simulation distortion.

[0029] In some other embodiments, the process of generating an underground cable distribution model using reverse software includes: Step 1: Complete the material library construction—Based on the electromagnetic properties of various substances in the actual underground environment, define the dielectric constant, conductivity, relative permeability, and magnetic loss of key materials, with the specific configuration as follows: Table 1. Parameters related to the underground cable distribution model The second step is 3D modeling of the complex underground scene. Specifically, a 1m×1m×1m underground scene is constructed to recreate the realistic distribution of urban underground soil, interference objects, and live cables. The specific steps are as follows: Setting up a surface air layer: An air layer is constructed in the range of z=0~0.15m to simulate the distance between the GPR antenna and the ground surface, avoiding signal reflection distortion caused by direct contact between the antenna and the soil; Setting up a non-uniform soil layer: Two layers are constructed to simulate soil stratification. The upper transition soil (z=0.15~0.6m): The Peplinski model (sand content 0.7%, clay content 0.3%, density 2.0g / cm³, reference frequency 2.66GHz, volumetric water content 0.25) is used to generate the non-uniform soil my_soil; then, a soil region with x=0~1m, y=0~1m, and z=0.15~0.6m is generated, and a surface roughness of 0.05~0.15m is added to simulate surface undulation; The lower foundation soil (z=0.6~1m): This is generated by adding roughness... Roughness is used to simulate the non-uniformity of deep soil. Irregular rocks are set up: five rocks at different locations are constructed within the soil layer to avoid unrealistic reflections caused by regular shapes. The material is irregular_stone, and surface roughness is added. The other four rocks (big_rock1~big_rock4) are modeled similarly according to the file coordinates. Air cavities are set up: spherical cavities with a radius of 0.05m are constructed at depths of x=0.75m, y=0.6m, and z=0.9m to simulate electromagnetic signal abrupt changes caused by underground pores. Other pipelines are set up: two types of non-cable pipelines are constructed to simulate pipeline intersections. The concrete underground pipe extends along the y-axis at x=0.7m, z=0.7m, and a radius of 0.07m, with the material concrete_underground_pipe. The plastic pipe extends along the y-axis at x=0.15m, z=0.35m, and a radius of 0.03m, with the material plastic_pipe. The other two plastic pipes are modeled similarly according to the simulation file coordinates. The cable distribution model formed after modeling is as follows. Figure 3 As shown, different cable distribution models can be formed by changing the coordinates of the cable distribution location and the coordinates of other interfering targets, thereby constructing a cable distribution model set. Specifically, by setting randomly generated pipes of various sizes and shapes, the non-uniform medium content of the soil and the size and location of the interfering objects are modified on this basis to achieve data augmentation strategies.

[0030] The third step, to simulate the unique electromagnetic characteristics of a "live cable," involves injecting a power frequency signal into the cable core through excitation. The specific steps are as follows: Define the live excitation waveform: Generate a 50Hz sine wave with an amplitude of 220V using a command to simulate the power frequency signal of an actual live cable. At this time, the cable insulation layer will generate periodic oscillation textures due to electromagnetic induction, becoming a key feature for subsequent model identification of the "live cable." Load the excitation signal into the cable core: The sine wave is transmitted through a 50Ω characteristic impedance transmission line to positions x=0.5m, y=0, and z=0.5m of the cable core, ensuring uniform signal transmission along the cable's extension direction. Mark the grounding boundaries at both ends of the cable's metal sheath using the #edge command to prevent the excitation signal from leaking into the soil and causing simulation distortion.

[0031] Step S102: Perform ultrasonic B-scan on the underground cable distribution model set using simulation software to obtain the scan dataset.

[0032] The scan dataset includes scan images, specifically scan maps obtained after ultrasonic scanning of underground surfaces, simulated by simulation software. Different B-scans are performed on different underground cable distribution models to obtain different mapping images, thus forming the scan dataset. The "B" in ultrasonic B-scan comes from the English word "brightness." The scan images are displayed in two dimensions; the screen shows a cross-section parallel to the direction of sound propagation and perpendicular to the measurement surface of the workpiece. The brightness information is determined by calculating the intensity of the reflected ultrasonic waves.

[0033] The following is an exemplary illustration of the specific process for constructing a scanned dataset: Step 1: Configure GPR (Ground Penetrating Radar) detection parameters. Specifically, set the detection parameters to be consistent with the actual GPR equipment, and collect B-scan data that includes "electrified features + complex interference". The parameter configuration is as follows: Antenna and waveform settings: The transmitting antenna is set to z-axis polarization, with positions x=0.05m, y=0.5m, z=0.215m, transmitting at 4e8 frequency, Ricker wavelet, center frequency 400MHz, balancing detection depth and resolution; the receiving antenna: #rx:0.070.50.215, 0.02m distance from the transmitting antenna, synchronously receiving reflected signals; Scanning and Time Window: The antenna moves 0.01m along the x-axis in one step, covering the range of x=0~1m; the time window is set to 1e-7, 100ns, corresponding to a maximum detection depth of about 2.5m, fully covering cable burial depth and deep interference; Boundary and Resolution: 10 perfectly matched layers are set for boundary conditions to absorb boundary reflection signals and avoid interference; The spatial resolution was set to 0.005, and the x / y / z axis step size was 5mm to ensure that the details of the cable's multi-layer structure were captured. B-scan data generation: One A-scan is acquired for each step of movement, accumulating to 100 A-scans, which are then stitched together in the x-axis movement order to form one B-scan image. A-scan comes from the English word Amplitude, meaning the amplitude. In other words, the horizontal axis of the display represents the propagation time or distance of the ultrasonic wave in the tested material, and the vertical axis represents the amplitude of the reflected ultrasonic wave.

[0034] The second step is to construct the GPR dataset. Specifically, 1,000 B-scan images simulating real underground scenarios are generated using script files. Spatial augmentation is used to expand the diversity of cable positions and sizes through random rotation and scaling. Physical augmentation is used to randomly adjust the antenna frequency and movement position of the simulated data based on physical parameters, simulating the dynamic changes of signals and different underground conditions in real detection, narrowing the gap between simulated and real data, and laying the foundation for improving the model's generalization ability.

[0035] Step S103: The scanned dataset is deeply fused using the trained data processing model to obtain a temporal fusion feature set.

[0036] Specifically, for each scan number in the scan dataset, temporal fusion is performed in the depth direction to extract cumulative evolutionary information from shallow to deep. This depth-oriented evolutionary information reveals the unique characteristics that distinguish the cable from other interference targets, such as the peak position of reflection intensity and attenuation rate, thereby achieving accurate identification of the cable signal. Deep fusion in the depth direction can be implemented using different models with corresponding functions, such as convolutional neural network models or long short-term memory models. Specific models and data deep fusion processes are detailed in other embodiments and will not be repeated here.

[0037] Step S104: Train the detection model using the temporal fusion feature set until the training is complete and the detection model is obtained.

[0038] Specifically, one temporal fusion feature from the temporal fusion feature set is input into the detection model one by one. The detection model outputs the predicted position of the cable. The predicted position is compared with the actual cable position marked in the training set to obtain the loss value. The loss value is backpropagated and gradient descent is used to search for the optimal weight parameters of the detection model for this time, so that the model trained in this generation reduces the loss value to the minimum. This process is called completing one training. The above process is repeated to iteratively train the detection model using different temporal fusion features from the temporal fusion feature set until the loss value of the detection model continuously converges to an interval less than a preset threshold. The detection model is then considered to have completed training.

[0039] The scan dataset used for training contains labels to indicate actual parameters related to the cable. The process of adding these labels is as follows: Data annotation needs to accurately describe the cable's location, physical characteristics, and temporal relationships. A three-level annotation system is adopted: basic annotation marks the cable's position in the image with bounding boxes, in coordinate format (x1, y1, x2, y2), and the category is uniformly "cable" to ensure that the model understands the cable's spatial range; feature annotation distinguishes two types of key regions through binary masks. In the strong reflection region mask M_high of the metal core, the pixel where the reflection peak is located is marked as 1, and the background pixel is marked as 0, corresponding to the high-frequency strong reflection characteristics of the cable's metal core; in the oscillating texture region mask M_low of the insulation layer, the pixel where the periodic stripes are located is marked as 1, and the background pixel is marked as 0, corresponding to the low-frequency oscillation characteristics of the insulation layer. Through this type of annotation, the cable's physical characteristics are transformed into supervisory signals that the model can learn.

[0040] Deep fusion focuses on the core characteristic of the "time-depth" correlation in GPRB scans. Since the vertical axis of a B-scan directly corresponds to the electromagnetic wave propagation time, annotation focuses on the evolution of cable features along the vertical axis. In a single B-scan, eight consecutive feature intervals are divided along the depth direction. Key information about the cable reflection features is annotated for each interval, including the local morphology of the cable hyperbola within that interval, the specific coordinates of the peak reflection intensity, and the feature relationships between adjacent depth intervals. This annotation provides temporal supervision along the depth direction—allowing the module to understand the continuous change logic of cable reflection features in the depth dimension. This helps the model learn and capture the inherent evolution of cable reflection signals along the depth direction, achieving accurate modeling of cable features in B-scans.

[0041] Optionally, after labeling the images with Labelme, a wavelet thresholding denoising algorithm is used to process the data. A dB4 wavelet basis is selected, and adaptive threshold calculation is used to suppress high-frequency noise caused by soil clutter, preserving the effective signal from cable reflections and avoiding noise interference with feature extraction. Subsequently, pixel value normalization is performed, transforming the pixel values ​​of all images to the [0,1] interval through linear mapping to eliminate amplitude differences caused by different detection devices and gain settings, ensuring consistent data distribution. Step S105: Input the actual detection scanning data into the data processing model to obtain temporal fusion features, and input the temporal fusion features into the detection model to obtain the detection results.

[0042] This can be understood as the data processing model being embedded in the front end of the detection model. During use, the actual scanning data obtained through ground-penetrating radar scanning is first input into the data processing model for data processing to obtain time-series fusion features before being input into the detection model, thereby enabling the detection model to have better recognition accuracy.

[0043] In some embodiments, such as Figure 4 As shown, Figure 1 The specific steps of step S103 include: Step S301: Process the data in the scanned dataset using physical constraints to obtain the filtered scanned dataset.

[0044] This can be understood as filtering and processing the data in the scan dataset based on the prior knowledge of the cable's special physical characteristics, obtaining filtered data that better reflects the cable's characteristics, and constructing a filtered scan dataset. This filtering can be done, for example, by filtering the cable data based on the characteristic that the cable data decreases with depth, or by processing it based on the cable's electromagnetic frequency characteristics.

[0045] In some other embodiments, the cable exhibits two typical characteristics in the GPR image: high-frequency strong reflection and low-frequency oscillation. The physical mechanism module separates and dynamically fuses these two characteristics, allowing the model to focus on the cable's unique electromagnetic response and suppress soil clutter and interference from other pipelines. Therefore, it is necessary to process the scan data in the high-frequency and low-frequency ranges. The processing procedure is illustrated below.

[0046] The first step is to perform image recognition on the scanned images in the scanned dataset to obtain high-frequency and low-frequency data.

[0047] Specifically, a visual recognition model is used to obtain high-frequency and low-frequency data based on the waveform characteristics of the cable at high and low frequencies. Optionally, the visual recognition model includes two channels, which are used to identify and extract high-frequency and low-frequency data respectively. The high-frequency branch extracts the strong reflection characteristics of the metal core. The high conductivity of the metal core causes electromagnetic waves to be strongly reflected on its surface, which is represented as "high amplitude spikes" in the GPR image, corresponding to high-frequency edge features. The low-frequency branch is used to capture the periodic oscillating texture of the insulation layer. The difference in dielectric constant between the insulation layer and the soil causes electromagnetic waves to be reflected multiple times inside the insulation layer, forming "periodic stripes", corresponding to low-frequency features.

[0048] The second step involves convolving the edge detection operator with the high-frequency data to obtain a high-frequency feature map, and then processing the low-frequency data through a dilated convolutional layer to obtain a low-frequency feature map.

[0049] Specifically, the high-frequency branch first reduces the number of channels in the backbone network's output feature map to 128 through 1×1 convolutions, reducing computational cost while achieving feature dimensionality reduction; then it connects to a 3×3 convolutional layer, with the convolution kernel employing an edge detection operator (…). By calculating pixel gradients to enhance edge changes in the reflected signal and highlight the location information of the metal core, the feature extraction formula is as follows: in Represents high-frequency feature maps, The backbone network outputs feature maps, and the LeakyReLU activation function is used to introduce nonlinearity to enhance the model's expressive power.

[0050] The low-frequency branch first reduces the number of channels to 128 using a 1×1 convolution, then connects to a dilated convolution layer. Dilated convolution expands the receptive field to 7×7 by inserting zeros into the kernel, capturing a wider range of pixel relationships and effectively extracting the periodic texture of the insulating layer. The feature extraction formula is as follows: in, Represents low-frequency feature maps. This represents dilated convolution.

[0051] The third step is to use an attention mechanism to perform weighted fusion of high-frequency feature maps and low-frequency feature maps to obtain the filtered scan dataset.

[0052] Specifically, to dynamically balance the contributions of high-frequency and low-frequency features, a weighting mechanism is introduced. Max pooling is performed on the high-frequency feature map to extract the maximum reflection intensity, reflecting the salience of the metal core; average pooling is performed on the low-frequency feature map to calculate the average texture energy, reflecting the salience of the insulating layer; the two pooling results are concatenated (⊕), the number of channels is compressed through 1×1 convolution and nonlinearity is introduced, and finally, the weights are generated by normalization using a Sigmoid activation function. ( ∈[0,1]), the weight calculation formula is as follows in The function is the Sigmoid function. Based on attention weights, high-frequency and low-frequency features are fused; the fusion formula is as follows: in, A higher value indicates that high-frequency features contribute more to recognition, while a lower value indicates that low-frequency features are more important, enabling adaptive feature selection in different scenarios. This indicates filtering the scanned data in the scanned dataset.

[0053] Step S302: Convert each training data in the filtered scan dataset into a pixel sequence in the depth direction.

[0054] This can be understood as follows: the core of the depth fusion module designed for B-scan images is to treat the signal distribution pattern in the depth direction as an "intrinsic temporal sequence," and to enhance the recognizability of cable features by modeling this temporal correlation. A B-scan image is essentially a two-dimensional data matrix of "lateral position-depth" (denoted as...). ,in The number of pixels in the depth direction. (where the horizontal distance is in pixels), for each horizontal position ( ), its depth direction pixel sequence ( For the first Column, No. The pixel values ​​in the depth sequence (representing the evolution of cable reflection signals with depth) reveal a strong reflection peak when the lateral position passes directly above the cable; this peak shifts to deeper depths as the lateral position moves to either side. Traditional YOLO models only extract local spatial features, ignoring this correlation in the depth sequence, making it difficult to capture the complete hyperbolic shape. This is especially problematic when signal attenuation or noise interference causes missed detections. The following describes a method to transform the traditional YOLO model into one that can obtain the pixel sequence along the depth direction.

[0055] Feature structure decomposition transforms the feature map output by the backbone network into a deep sequence format. Let the feature map at a certain scale output by the YOLO backbone network be... ( For the number of channels, , (Corresponding to feature dimensions in the depth and lateral directions respectively), the module first categorizes by lateral position. To split, we get A single "deep feature sequence": Each This multi-channel feature map, encompassing the lateral location at all depths, allows shallow channels to capture edge details of the reflected signal and deeper channels to encode the overall trend in reflection intensity. This decomposition transforms the two-dimensional feature map into a series of one-dimensional depth sequences, laying the foundation for subsequent temporal modeling.

[0056] Step S303: Process the pixel sequence in the depth direction using a unidirectional long short-term memory network to obtain multiple hidden layer parameters.

[0057] Specifically, deep sequence modeling uses recurrent neural networks to capture the evolutionary patterns along the depth direction. Since electromagnetic waves propagate unidirectionally underground, a unidirectional LSTM is used to model the sequence. ,in For LSTM in the first A deep hidden state, The initial zero vector is used. The LSTM's gating mechanism can automatically filter key information—when the depth... When a strong reflection signal appears at a certain point, the input gate strengthens the weight of that feature, while the forget gate weakens the interference from shallow, non-critical features, ultimately making... Includes from light to dark ( The accumulated evolution information of the reflection intensity, such as the peak position and attenuation rate, is obtained. After LSTM processing, the depth sequence at each lateral position is transformed into a hidden state sequence. It fully preserves the feature associations in the depth direction.

[0058] Step S304: The parameters of each hidden layer are weighted and fused through an attention mechanism to obtain a temporal fusion feature set.

[0059] This can be understood as follows: the most crucial aspect of deep fusion mechanism recognition is depth features. Different depths contribute significantly to cable recognition. The module introduces attention weights to the hidden state sequence. The weighted fusion process involves the following steps: First, an importance score for each hidden state is calculated using a multilayer perceptron (MLP). The MLP consists of two fully connected layers (with LeakyReLU activation in the middle layer), outputting a one-dimensional score; subsequently, the score is normalized into weights using the Softmax function. Ensure the sum of the weights is 1; finally, perform weighted fusion to obtain the deep aggregated feature at this horizontal position: All horizontal positions Reorganized into spatial feature map (dimension is) To supplement the spatial correlation in the lateral direction, a 1×H convolutional layer (kernel size 1×H, stride 1) is used. Process it to make its dimensions match those of the original feature map. Match (converted to) ), and then with Element-by-element addition and fusion are performed to obtain the final "spatial-depth" joint feature map. It retains the spatial details of the original features while injecting temporal correlation information of the depth sequence, enabling the output features to depict the cable's positional distribution in the lateral direction and reflect its morphological changes in the depth direction.

[0060] In some other embodiments, such as Figure 5 As shown, the training methods for the detection model include: The process involves generating a series of virtual underground cable models and corresponding ground-penetrating radar (GPRmax) scan data using GPRmax software. This step includes setting the GPR and medium parameters, adding underground stratification, non-cable pipes, irregular rocks, and other interfering objects, and adding excitation signals to simulate live cables.

[0061] The scanned data is input into the GPM-YOLO model for processing. The basic spatial features of the scanned data are extracted through the GPM-YOLO model. Physical feature detection is performed on the scanned data, and temporal fusion is performed on the scanned data in the depth direction. The temporal fusion features are input into the feature pyramid so that the temporal fusion can take into account different scale ranges. Finally, the training set is obtained through classification and regression. The actual parameters related to the cable are labeled in advance in the training set.

[0062] It should be noted that the above two steps can be completed in batch processing, that is, after the GPRmax software simulation software generates a batch of scan data, the GPM-YOLO model processes these scan data in batches; the above two steps can also be processed in parallel in a stream-out manner, that is, after the GPRmax software simulation software generates a scan data, the scan data is input into the GPM-YOLO model for processing, while the GPRmax software simulation software continues to generate the next scan data.

[0063] After generating the training set, the detection model is tested using the data in the training set. After training is completed, the detection effect of the detection model is verified using the validation set, and the detection results are displayed.

[0064] In some embodiments, such as Figure 6 As shown, the process of completing the exclusive configuration for a specific target by adding a physical attention mechanism module and a multi-depth temporal fusion module to the GPM-YOLO model includes: inputting the scanning data image of GPR (Ground Penetrating Radar) into the GPM-YOLO model, extracting general basic features through the Backbone architecture, detecting physical features through the PAM clustering algorithm, performing data fusion through the Neck layer, performing deep feature fusion through MDTFM (Multi-task Delayed Feedback Model), and outputting the cable detection result through the Head detection head. This detection result is not the final detection result and needs to be input into the trained detection model to obtain the final detection result.

[0065] This invention also provides an underground cable detection system based on the fusion of real-world simulation and time-series data, such as... Figure 7As shown, the cable detection system includes: a simulation module 100, a fusion module 200, a training module 300, and a detection module 400. The simulation module 100 is used to generate an underground cable distribution model set using simulation software, and also to perform ultrasonic B-scans on the underground cable distribution model set using simulation software to obtain a scan dataset, which includes scan images. The fusion module 200 is used to perform deep fusion of the scan dataset using a trained data processing model to obtain a temporal fusion feature set. The training module 300 is used to train the detection model using the temporal fusion feature set until training is complete and a detection model is obtained. The detection module 400 is used to input actual detection scan data into the data processing model to obtain temporal fusion features, and input the temporal fusion features into the detection model to obtain the detection result.

[0066] In some embodiments, such as Figure 7 As shown, the fusion module 200 is also used to process the data in the scan dataset through physical constraints to obtain a filtered scan dataset; it is also used to convert each training data in the filtered scan dataset into a pixel sequence in the depth direction; it is also used to process the pixel sequence in the depth direction through a unidirectional long short-term memory network to obtain multiple hidden layer parameters; and it is also used to perform weighted fusion of each hidden layer parameter through an attention mechanism to obtain a temporal fusion feature set.

[0067] In some embodiments, such as Figure 7 As shown, the fusion module 200 is also used to calculate the importance score of each hidden layer parameter through the multilayer perception mechanism, and obtain the importance weight after normalizing the importance score; it is also used to use the importance weight as the weight of the corresponding hidden layer parameter to perform weighted fusion of each hidden layer parameter.

[0068] In some embodiments, such as Figure 7 As shown, the fusion module 200 is also used to perform image recognition on the scanned images in the scanned dataset to obtain high-frequency data and low-frequency data; it is also used to perform convolution calculation with the high-frequency data through the edge detection operator to obtain a high-frequency feature map, and to process the low-frequency data through the dilated convolutional layer to obtain a low-frequency feature map; it is also used to perform weighted fusion of the high-frequency feature map and the low-frequency feature map through the attention mechanism to obtain a filtered scanned dataset.

[0069] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for detecting underground cables based on the fusion of realistic simulation and time series, characterized in that, The underground cable detection method includes: A set of underground cable distribution models was created using simulation software; The underground cable distribution model set was simulated by the simulation software to perform ultrasonic B-scan, and a scan dataset was obtained, which included scan images. The trained data processing model is used to perform deep fusion on the scanned dataset to obtain a temporal fusion feature set. The detection model is trained using the time-series fusion feature set until training is complete and the detection model is obtained. The actual detection scanning data is input into the data processing model to obtain temporal fusion features, and the temporal fusion features are input into the detection model to obtain the detection result.

2. The underground cable detection method according to claim 1, characterized in that, The process of deep fusing the scanned dataset using the trained data processing model to obtain a temporal fusion feature set includes: The data in the scanned dataset is processed by physical constraints to obtain a filtered scanned dataset; Each training data point in the filtered scan dataset is converted into a pixel sequence along the depth direction; Multiple hidden layer parameters are obtained by processing the pixel sequence in the depth direction through a unidirectional long short-term memory network; The parameters of each hidden layer are weighted and fused using an attention mechanism to obtain the temporal fusion feature set.

3. The underground cable detection method according to claim 2, characterized in that, The step of weighted fusing the parameters of each hidden layer through an attention mechanism to obtain the temporal fusion feature set includes: The importance score of each hidden layer parameter is calculated through a multilayer perceptron mechanism, and the importance weight is obtained by normalizing the importance score. The importance weights are used as the weights of the corresponding hidden layer parameters to perform weighted fusion of each hidden layer parameter.

4. The underground cable detection method according to claim 2, characterized in that, The process of processing the data in the scanned dataset through physical constraints to obtain the filtered scanned dataset includes: Image recognition is performed on the scanned images in the scanned dataset to obtain high-frequency data and low-frequency data; A high-frequency feature map is obtained by convolving the high-frequency data with the edge detection operator, and a low-frequency feature map is obtained by processing the low-frequency data through a dilated convolutional layer. The high-frequency feature map and the low-frequency feature map are weighted and fused using an attention mechanism to obtain the filtered scan dataset.

5. The underground cable detection method according to claim 4, characterized in that, The process of weightedly fusing the high-frequency feature map and the low-frequency feature map using an attention mechanism to obtain the filtered scanning dataset includes: Max pooling is performed on the high-frequency feature map to extract the maximum reflection intensity, which reflects the significance of the metal core; Average pooling is performed on the low-frequency feature map to calculate the average texture energy, which reflects the saliency of the insulating layer; The maximum reflection intensity and the average texture energy are concatenated by channels. The result of the channel concatenation is compressed into one dimension by a one-dimensional convolutional layer, and the weights are obtained by normalization. The high-frequency feature map and the low-frequency feature map are weighted and fused using the weights to obtain the filtered scan dataset.

6. The underground cable detection method according to claim 1, characterized in that, The set of underground cable distribution models generated through simulation software includes: Construct a virtual model of underground cables, and set up cable models within the underground cable model; The underground cable model includes a non-cable conduit model, an irregular rock model, and an air cavity model. The cable model is stimulated by injecting a power frequency signal.

7. An underground cable detection system based on the fusion of realistic simulation and time sequence, characterized in that, The underground cable detection system includes: The simulation module is used to generate a set of underground cable distribution models through simulation software, and is also used to perform ultrasonic B-scans on the underground cable distribution model set through the simulation software to obtain a scan dataset, which includes scan images. The fusion module is used to perform deep fusion of the scanned dataset through the trained data processing model to obtain a temporal fusion feature set; The training module is used to train the detection model using the temporal fusion feature set until the training is complete and the detection model is obtained. The detection module is used to input the actual detection scan data into the data processing model to obtain temporal fusion features, and input the temporal fusion features into the detection model to obtain the detection result.

8. The underground cable detection system according to claim 7, characterized in that, The fusion module is further configured to process the data in the scan dataset through physical constraints to obtain a filtered scan dataset; to convert each training data in the filtered scan dataset into a pixel sequence in the depth direction; to process the pixel sequence in the depth direction through a unidirectional long short-term memory network to obtain multiple hidden layer parameters; and to perform weighted fusion of each hidden layer parameter through an attention mechanism to obtain the temporal fusion feature set.

9. The underground cable detection system according to claim 8, characterized in that, The fusion module is also used to calculate the importance score of each hidden layer parameter through a multilayer perceptron mechanism, and to obtain the importance weight after normalizing the importance score; It is also used to perform weighted fusion of each hidden layer parameter by using the importance weight as the weight of the corresponding hidden layer parameter.

10. The underground cable detection system according to claim 8, characterized in that, The fusion module is further configured to perform image recognition on the scanned images in the scanned dataset to obtain high-frequency data and low-frequency data; to perform convolution calculation with the high-frequency data using an edge detection operator to obtain a high-frequency feature map; to process the low-frequency data using a dilated convolutional layer to obtain a low-frequency feature map; and to perform weighted fusion of the high-frequency feature map and the low-frequency feature map using an attention mechanism to obtain the filtered scanned dataset.

Images