A ground penetrating radar underground anomaly cooperative risk assessment method, system and device

Abstract

This invention provides a method, system, and device for collaborative risk assessment of underground anomalies using ground-penetrating radar. The method includes: acquiring detection data and environmental context information of the detection area; constructing a multi-task learning model, which includes a shared feature extraction network and multiple parallel collaborative task branches, including at least an anomaly detection and classification branch, a pixel-level segmentation and quantization branch, and a risk assessment branch; training the multi-task learning model using a multi-task labeled dataset and an adaptive weighted total loss function to obtain an inference model; inputting the detection data and environmental context information into the inference model, and synchronously outputting inference results from the multiple task branches through a single forward propagation; and integrating the inference results from multiple task branches to generate a detection and diagnostic report. This method achieves simultaneous detection, quantification, and risk assessment of underground anomalies in a single forward inference, streamlining the entire automated process.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of geophysical exploration, artificial intelligence and underground engineering safety, and specifically relates to a method, system and equipment for collaborative risk assessment of underground anomalies using ground penetrating radar. Background Technology

[0002] With the continuous expansion of urban underground engineering construction, the operational safety of infrastructure such as subways, tunnels, urban roads, and integrated utility tunnels is directly related to urban public safety. Ground penetrating radar (GPR), as an efficient and non-destructive geophysical detection method, is widely used in the detection of underground cavities, voids, loose bodies, and other defects, as well as the identification of concealed structures such as pipelines and reinforcing bars.

[0003] There are three major pain points in the current interpretation and application of GPR data:

[0004] First, traditional manual interpretation is extremely inefficient and highly subjective. Manual interpretation of a single 100-meter-level profile takes 2-4 hours, heavily relies on the professional experience and skills of technicians, and the interpretation results are highly subjective with inconsistent evaluation standards, which cannot meet the engineering needs of large-scale, high-frequency city-level exploration.

[0005] Second, existing AI technologies suffer from fragmented functions, leading to the accumulation of errors at each stage. Most publicly available AI solutions are single-task models, capable of only single target detection or semantic segmentation. They generally employ a sequential processing flow of "detect first, then segment, then evaluate," where errors from earlier stages are amplified and propagated at subsequent stages, making it impossible to achieve end-to-end integrated interpretation from raw data to output results.

[0006] Third, risk assessment is detached from industry standards and cannot directly support engineering decisions. Existing technologies cannot automatically match the mandatory requirements and risk assessment systems of industry standards such as the "Technical Specification for Detection of Urban Underground Pathogens" DB11 / T 1399-2017 and the "Technical Standard for Comprehensive Detection and Risk Assessment of Urban Underground Pathogens" JGJ / T 437-2018. They cannot complete standardized risk level classification and treatment recommendations, and the interpretation results are difficult to directly support engineering decisions and closed-loop management. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a ground-penetrating radar method, system, and equipment for collaborative risk assessment of underground anomalies. This method enables simultaneous detection, classification, precise quantification, and standardized risk assessment of underground anomalies in a single forward inference step, thus streamlining the entire automated process.

[0008] The technical solution of the present invention is as follows:

[0009] A method for collaborative risk assessment of underground anomalies using ground-penetrating radar includes the following steps:

[0010] S1: Acquire detection data and environmental context information of the detection area;

[0011] S2: Construct a multi-task learning model, which includes a shared feature extraction network and multiple parallel and collaborative task branches. The multiple task branches include at least an anomaly detection and classification branch, a pixel-level segmentation and quantization branch, and a risk assessment branch.

[0012] S3: Train the multi-task learning model using a multi-task labeled dataset and an adaptive weighted total loss function to obtain the inference model;

[0013] S4: Input the detection data and environmental context information into the inference model, and through one forward propagation, the inference results are synchronously output by the multiple task branches;

[0014] S5: Integrate the inference results from multiple task branches to generate a detection and diagnostic report.

[0015] Furthermore, in the aforementioned ground-penetrating radar underground anomaly collaborative risk assessment method, in step S1:

[0016] The detection data is ground-penetrating radar B-Scan data; and / or,

[0017] After acquiring the detection data, the process further includes preprocessing the data and obtaining the equivalent electromagnetic wave velocity of the underground medium in the detection area through calibration, in order to establish a time-depth conversion mapping relationship between the radar two-way travel time and the actual burial depth; and / or,

[0018] The environmental context information includes at least one of the following: load conditions, geological conditions, and surrounding environmental attributes of the area where the abnormal target is located.

[0019] Furthermore, in the aforementioned ground-penetrating radar underground anomaly collaborative risk assessment method, the preprocessing includes at least one of DC offset removal, zero-point correction, background noise removal, bandpass filtering, time gain compensation, and data normalization.

[0020] Furthermore, in the ground-penetrating radar underground anomaly collaborative risk assessment method, the shared feature extraction network in step S2 adopts a deep convolutional neural network or a visual Transformer network pre-trained on a general image dataset. Through transfer learning, it is fine-tuned on the special dataset of the detection data, and the extracted shared features are adapted to the feature requirements of the multiple task branches.

[0021] Furthermore, in the aforementioned ground-penetrating radar underground anomaly collaborative risk assessment method, the multiple task branches in step S2 specifically include:

[0022] The anomaly detection and classification branch adopts an integrated attention mechanism target detection architecture to output the bounding boxes, category labels, and classification confidence scores of underground anomalies in the detection data.

[0023] The pixel-level segmentation and quantization branch uses a semantic segmentation network with an encoder-decoder structure to output pixel-level segmentation masks for anomalous targets and automatically calculate their geometric quantization parameters.

[0024] The risk assessment branch employs a lightweight multilayer perceptron network. Its inputs include at least the geometric quantization parameters, the category labels, and the environmental context information. Its outputs include the risk level and corresponding handling recommendations.

[0025] Furthermore, in the aforementioned ground-penetrating radar underground anomaly collaborative risk assessment method, the target detection architecture integrating the attention mechanism is any one of improved YOLOv8, YOLOv10, or RetinaNet; and / or,

[0026] The category labels include at least six categories: voids, cavities, loose bodies, water-rich bodies, pipelines, and reinforcing steel; and / or,

[0027] The semantic segmentation network is DeepLabv3+ or U-Net++; and / or,

[0028] The geometric quantification parameters include the top burial depth, bottom burial depth, vertical height, horizontal span, and projected area of ​​the anomalous target; and / or,

[0029] The calculation logic of the risk assessment branch embeds the risk assessment system of the "Technical Specification for Detection of Urban Underground Diseases" DB11 / T1399-2017 and / or the "Technical Standard for Comprehensive Detection and Risk Assessment of Urban Underground Diseases" JGJ / T 437-2018, automatically calculates the probability level of risk occurrence, the risk consequence level, and the comprehensive risk value, and classifies the risk level accordingly.

[0030] Furthermore, in the aforementioned ground-penetrating radar underground anomaly collaborative risk assessment method, each sample in the multi-task labeled dataset in step S3 contains a one-to-one quadruple label: anomaly detection box and category label, pixel-level segmentation mask label, geometric quantization parameter ground truth label, and risk occurrence probability level, risk consequence level, comprehensive risk value, and risk level ground truth label calculated according to industry standards; and / or,

[0031] The formula for the adaptive weighted total loss function in step S3 is:

[0032] ;

[0033] in, CIoU loss for anomaly detection and classification branches, The Dice loss is used for pixel-level segmentation and quantization branches. The cross-entropy loss for the risk assessment branch; , , These are adaptive weights for the losses of the three tasks, which can be dynamically adjusted during training.

[0034] Furthermore, in the ground-penetrating radar underground anomaly collaborative risk assessment method, step S5 also generates a visualized image with anomaly annotations and risk level rendering, as well as an anomaly result statistical table that conforms to industry standards.

[0035] The format and fields of the abnormal results statistics table match the requirements of Appendix G of the "Technical Standard for Comprehensive Detection and Risk Assessment of Urban Underground Diseases" JGJ / T 437-2018. The abnormal results statistics table shall at least include the fields of abnormality number, type, location, center coordinates, projected area, depth of impact, risk level, and treatment suggestion.

[0036] A ground-penetrating radar-based collaborative risk assessment system for underground anomalies includes:

[0037] The data acquisition module is used to acquire detection data and environmental context information of the detection area;

[0038] The model inference module has a built-in multi-task learning model, which includes a shared feature extraction network and multiple parallel and collaborative task branches. The multiple task branches include at least an anomaly detection and classification branch, a pixel-level segmentation and quantization branch, and a risk assessment branch. The model inference module is used to input the detection data and environmental context information into the multi-task learning model, and through one forward propagation, the multiple task branches synchronously output the inference results.

[0039] The report generation module is used to integrate the reasoning results of multiple task branches and generate a detection and diagnostic report.

[0040] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the ground-penetrating radar underground anomaly collaborative risk assessment method.

[0041] The beneficial effects of this invention are as follows:

[0042] This invention presents a collaborative risk assessment method for underground anomalies using ground-penetrating radar. By constructing an end-to-end model of "shared feature extraction network + multi-task parallel collaboration," it simultaneously completes the detection, classification, precise quantification, and standardized risk assessment of underground anomalies in a single forward propagation. This streamlines the entire automated process from "raw data to intelligent identification, precise quantification, and risk decision-making," eliminating the error accumulation problem of traditional sequential solutions. This method is adaptable to various underground engineering scenarios such as urban roads, subway tunnels, soft soil foundations, and integrated utility tunnels. The backbone network and task branches can be flexibly upgraded and replaced, demonstrating good technological evolution and scalability.

[0043] This ground-penetrating radar underground anomaly collaborative risk assessment method reduces the interpretation time of a single 100-meter-level profile from the traditional 2-4 hours to less than 1 second, improving efficiency by more than 1,000 times. It completely solves the industry pain points of low manual interpretation efficiency and inability to adapt to large-scale detection needs.

[0044] The ground-penetrating radar underground anomaly collaborative risk assessment method achieves a burial depth quantification error of ≤5cm through pixel-level segmentation and time-depth conversion coupled calculation; the risk assessment logic strictly conforms to industry standards and has a consistency of ≥92% with the rating results of senior experts, fundamentally eliminating the subjective arbitrariness of manual interpretation.

[0045] This ground-penetrating radar underground anomaly collaborative risk assessment method completes the entire process of detection, quantification, and assessment simultaneously through a single forward inference, completely avoiding the problem of error amplification at each stage in the traditional serial process, and achieving a high degree of consistency in the output information of each stage.

[0046] This ground-penetrating radar underground anomaly collaborative risk assessment method can directly generate detection results and reports that meet industry standards, realizing full-process automation from raw data to delivered results, greatly reducing the workload of technical personnel, and the results can be directly used for engineering decision-making and closed-loop management. Attached Figure Description

[0047] Figure 1 This is a flowchart of a ground-penetrating radar underground anomaly collaborative risk assessment method according to the present invention;

[0048] Figure 2 This is a flowchart of the collaborative risk assessment method for underground anomalies using ground-penetrating radar according to the present invention.

[0049] Figure 3 This is a schematic diagram of the multi-task learning model architecture of a ground-penetrating radar underground anomaly collaborative risk assessment method according to the present invention;

[0050] Figure 4 This is a flowchart of the risk assessment branch of a ground-penetrating radar underground anomaly collaborative risk assessment method according to the present invention;

[0051] Figure 5 This is a schematic diagram of the field engineering application of Embodiment 1 of the ground-penetrating radar underground anomaly collaborative risk assessment method of the present invention. Detailed Implementation

[0052] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The advantages and features of the present invention will become clearer from the following description and claims. It should be noted that the drawings are all in a very simplified form and use non-precise proportions, and are only used to facilitate and clarify the illustration of the embodiments of the present invention.

[0053] like Figure 1 and Figure 2 As shown in the figure, this embodiment provides a method for collaborative risk assessment of underground anomalies using ground penetrating radar, including the following steps: S1-S5. This method achieves "one-time feature extraction, multi-task sharing and reuse" by constructing an end-to-end multi-task learning model consisting of a "shared feature backbone network + multiple parallel and mutually collaborative task branches," thereby fundamentally avoiding the error accumulation problem of serial processes.

[0054] S1: Acquire detection data and environmental context information of the detection area.

[0055] The detection data refers to the raw B-Scan data collected in the detection area by ground-penetrating radar equipment. Environmental context information refers to background information related to the detection area that may affect the risk assessment results.

[0056] Furthermore, after acquiring the detection data, it undergoes preprocessing. The purpose of preprocessing is to suppress noise interference, enhance the effective signal, and unify the data scale, providing high-quality input for the subsequent multi-task learning model. Preprocessing includes at least one of the following: DC offset removal, zero-point correction, background noise removal, bandpass filtering, time gain compensation, and data normalization. DC offset removal is used to eliminate the DC component introduced by the radar system itself; zero-point correction is used to align the time starting point of the surface reflected waves; background noise removal is used to filter out continuous background interference in the horizontal direction; bandpass filtering is used to retain the effective signal frequency band and suppress high and low frequency noise; time gain compensation is used to compensate for the energy attenuation of electromagnetic waves during underground propagation; and data normalization is used to unify the signal amplitude to a preset range, improving the stability of model training.

[0057] Furthermore, this includes time-depth conversion calibration. Calibration is performed using data from targets or boreholes at known depths to obtain the equivalent electromagnetic wave velocity *v* of the subsurface medium in the detection area. Based on this equivalent electromagnetic wave velocity, a time-depth conversion mapping relationship is established between the radar's two-way travel time *t* and the actual burial depth *d*, following the formula: *d=(t×v) / 2*. This mapping relationship allows subsequent pixel-level segmentation and quantization branches to convert the pixel scale in image space into the geometric dimensions in physical space.

[0058] Environmental context information includes at least one of the following: load conditions, geological conditions, and surrounding environmental attributes of the area where the anomalous target is located. For example, load conditions include road load ratings; geological conditions include regional geological condition indicators; and surrounding environmental attributes include the horizontal distance between the anomalous target and nearby pipelines, surface deformation, importance of surrounding facilities, and population density in the area. This environmental context information is encoded and input into the model in the form of structured data, serving as an important basis for the compliance assessment of the risk assessment branch.

[0059] S2: Construct a multi-task learning model.

[0060] Construct an end-to-end multi-task learning model. For example... Figure 3 As shown, this multi-task learning model adopts an overall architecture of shared feature extraction network and parallel collaborative multiple task branches, and realizes information reuse and mutual promotion among multiple tasks by sharing underlying features.

[0061] The shared feature extraction network is a fundamental component of multi-task learning models, responsible for extracting general visual features from the preprocessed probe data. This network can employ a deep convolutional neural network (CNN) or a visual Transformer network pre-trained on a general image dataset (e.g., ImageNet), and then fine-tuned on a specialized probe dataset through transfer learning. The extracted shared features simultaneously adapt to the feature requirements of multiple task branches. The deep CNN can be any of the ResNet series (e.g., ResNet-50, ResNet-101) or the EfficientNet series (e.g., EfficientNet-B0 to EfficientNet-B7). The visual Transformer network can be selected from the Swin Transformer series. During fine-tuning, the network's initial weights are inherited from the pre-trained model, followed by several rounds of training on the GPR-specific dataset to adapt the network weights to the data distribution characteristics of GPR images. The extracted shared features are simultaneously fed into various downstream task branches to meet the differentiated feature requirements of different tasks.

[0062] The multiple task branches specifically include anomaly detection and classification, pixel-level segmentation and quantization, and risk assessment. These three branches work in parallel and collaboratively, sharing the output of the feature extraction network as their common input.

[0063] The anomaly detection and classification branch employs an integrated attention mechanism-based target detection architecture to output bounding boxes, category labels, and classification confidence scores for underground anomalies in the detection data. The integrated attention mechanism-based target detection architecture can be an improved version of YOLOv8, YOLOv10, or RetinaNet. Category labels include at least six types: cavities, voids, loose bodies, water-rich bodies, pipelines, and reinforcing steel, covering common types of defects and structures found in urban underground detection.

[0064] The pixel-level segmentation and quantization branch employs an encoder-decoder structured semantic segmentation network to output pixel-level segmentation masks for anomalous targets, achieving fine-grained recognition at the pixel level. This pixel-level segmentation and quantization branch shares underlying features with the aforementioned anomaly detection and classification branches, and achieves collaborative optimization through a mutual supervision learning mechanism—candidate regions provided by the detection branch can guide the segmentation branch to perform fine-grained segmentation at corresponding locations, while the masks output by the segmentation branch can be used to verify the accuracy of the detection boxes. The semantic segmentation network can be DeepLabv3+ or U-Net++.

[0065] After obtaining the pixel-level segmentation mask, the pixel-level segmentation and quantization branch automatically calculates the geometric quantization parameters of the anomalous target. Combined with the temporal-depth transformation mapping relationship established in S1, the pixel dimensions in image space are converted into geometric dimensions in physical space. The geometric quantization parameters include the top burial depth, bottom burial depth, vertical height, horizontal span, and projected area of ​​the anomalous target. The top burial depth is the vertical distance from the upper boundary of the anomalous target to the ground surface; the bottom burial depth is the vertical distance from the lower boundary of the anomalous target to the ground surface; the vertical height is the difference between the bottom and top burial depths; the horizontal span is the maximum horizontal spread length of the anomalous target; and the projected area is the projected coverage area of ​​the anomalous target on the horizontal plane.

[0066] like Figure 4 As shown, the risk assessment branch employs a lightweight Multi-Layer Perceptron (MLP) network. The input to this risk assessment branch includes at least three parts: geometric quantization parameters output from the pixel-level segmentation and quantization branch, category labels output from the anomaly detection and classification branch, and the aforementioned acquired environmental context information. The core feature of this risk assessment branch is that its computational logic embeds the risk assessment system of the "Technical Specification for Detection of Urban Underground Diseases" DB11 / T 1399-2017 and / or the "Technical Standard for Comprehensive Detection and Risk Assessment of Urban Underground Diseases" JGJ / T 437-2018.

[0067] The internal calculation logic of the risk assessment branch is as follows:

[0068] In one specific implementation, the risk assessment branch employs a lightweight 3-layer Multilayer Perceptron (MLP) network. Its inputs include the geometric quantization parameters, the category code obtained by one-hot encoding of the category label, and the environmental context information. After mapping through a fully connected network, the probability level P and the consequence level C of the risk are calculated respectively. The comprehensive risk value (range 1-25) is obtained by multiplying them by R=P×C. Based on the risk matrix of embedded industry standards, the risk is classified into risk levels I-V, and corresponding standardized handling recommendations are output.

[0069] Specifically, firstly, based on the category label and geometric quantification parameters (such as burial depth, height, area, etc.) of the abnormal target, and referring to the evaluation criteria stipulated in industry standards, the probability level P of risk occurrence is automatically determined. The assessment of the probability of risk occurrence comprehensively considers factors such as the type of abnormal target (different types of hazards, such as cavities and voids, have different degrees of harm), scale (the larger the area and the higher the height, the greater the probability), and burial depth (shallow-buried defects are more likely to induce collapse). Secondly, based on environmental context information (such as road load level, importance of surrounding facilities, and regional population density, etc.), and referring to the consequence severity assessment criteria stipulated in industry standards, the consequence level C of risk is automatically determined. Finally, combining the probability level of risk occurrence and the consequence level of risk, the comprehensive risk value R=P×C is calculated, and the risk level is classified (such as Level I - low risk, Level II - relatively low risk, Level III - medium risk, Level IV - high risk, Level V - extremely high risk), and corresponding disposal recommendations are output (such as "recommend immediate enclosure and excavation verification", "recommend borehole verification", "recommend inclusion in the regular monitoring plan", etc.).

[0070] S3: Train a multi-task learning model using a dataset containing multi-task annotations and an adaptive weighted total loss function to obtain an inference model.

[0071] A multi-task labeled dataset is constructed, in which each sample contains a one-to-one quadruple annotation: anomaly detection bounding box and category label, pixel-level segmentation mask label, geometric quantization parameter ground truth label, and risk probability level, risk consequence level, comprehensive risk value, and risk level ground truth label calculated according to industry standards. This quadruple annotation covers all supervision signals from the three task branches and forms the data foundation for end-to-end multi-task joint training of the model. For example, the dataset is divided into training, validation, and test sets according to a preset ratio (e.g., 7:2:1).

[0072] An adaptive weighted total loss function is used to coordinate the joint optimization of the three task branches. A key challenge in multi-task learning is that the magnitudes and convergence speeds of the loss functions differ across tasks, and fixed weights may lead to one task dominating training or another failing to converge effectively. Therefore, an adaptive weighting strategy is employed.

[0073] The formula for the adaptive weighted total loss function is: .

[0074] in, For the loss of the anomaly detection and classification branches, CIoU (Complete Intersectionover Union) loss is adopted. This loss comprehensively considers the intersection-union ratio, center point distance and aspect ratio consistency between the predicted box and the ground truth box, and can effectively measure the accuracy of object detection. For the loss of pixel-level segmentation and quantization branches, Dice loss is adopted. This loss directly optimizes the overlap between the predicted mask and the real mask and has strong robustness to the class imbalance problem. For the loss of the risk assessment branch, cross-entropy loss is adopted, which is suitable for multi-classification tasks with risk levels. , , These are adaptive weights for the losses of the three tasks. The initial weight values ​​can be set to 0.4, 0.4, and 0.2, and are dynamically adjusted during training based on the convergence of each branch. For example, strategies such as uncertainty weighting or gradient magnitude balancing can be used for adaptive adjustment to ensure that the three tasks converge to the optimal state at a relatively balanced speed.

[0075] During training, training set samples are input into the multi-task learning model in batches. The output of each branch is calculated through forward propagation, and the total loss is calculated by combining the four-fold annotation. The model parameters are updated through backpropagation, iterating until the model's performance metrics converge on the validation set. Finally, the model performance is evaluated on the test set, and the optimal model parameters are saved as the inference model.

[0076] S4: Input the probe data and environmental context information into the inference model, and output the inference results synchronously from multiple task branches through one forward propagation.

[0077] This step is the end-to-end parallel inference stage. Input data includes the probe data preprocessed by S1 and the environmental context information of the corresponding probe region. The inference model, through one forward propagation, feeds the features extracted by the shared feature extraction network in parallel to three task branches. The three branches simultaneously complete their respective inference calculations and output: bounding boxes, category labels, and classification confidence scores from the anomaly detection and classification branch; pixel-level segmentation masks and geometric quantization parameters from the pixel-level segmentation and quantization branch; and risk level and handling recommendations from the risk assessment branch.

[0078] The entire reasoning process can be completed with only one forward propagation, eliminating the need for serial waiting and intermediate result passing in traditional schemes.

[0079] S5: Integrates the reasoning results from multiple task branches to generate visual images with anomaly annotations and risk level rendering, statistical tables of abnormal results that conform to industry standards, and standardized detection and diagnostic reports.

[0080] The format and fields of the abnormal results statistics table should match the requirements of Appendix G of the "Technical Standard for Comprehensive Detection and Risk Assessment of Urban Underground Diseases" JGJ / T 437-2018, and should at least include the fields of abnormality number, type, location, center coordinates, projected area, depth of impact, risk level, and treatment recommendations.

[0081] This step integrates the reasoning results synchronously output in step S4 to form outputs oriented towards engineering applications, which can be directly used for engineering decision-making, construction briefings, and archiving management.

[0082] The above method constructs an end-to-end deep learning model of "shared feature extraction network + multi-task parallel collaboration" to simultaneously complete the detection, classification, accurate quantification and standard-driven risk assessment of underground anomalies in one forward propagation. It opens up a fully automated process from "raw data to intelligent identification, to accurate quantification and to risk decision-making", eliminating the error accumulation problem of traditional serial solutions.

[0083] This embodiment also provides a ground-penetrating radar underground anomaly collaborative risk assessment system, including a data acquisition module, a model inference module, and a report generation module.

[0084] The data acquisition module is used to acquire detection data and environmental context information of the detection area. This module can communicate with ground-penetrating radar equipment to directly receive collected B-Scan data or read historical data from data storage devices. Environmental context information can be manually entered by the user or automatically obtained from external data sources such as geographic information systems and municipal pipeline network databases.

[0085] The model inference module incorporates a multi-task learning model. This model includes a shared feature extraction network and multiple parallel, collaborative task branches, each comprising at least an anomaly detection and classification branch, a pixel-level segmentation and quantization branch, and a risk assessment branch. The model inference module receives probe data and environmental context information from the data acquisition module. Through a single forward propagation, multiple task branches simultaneously output inference results. These inference results include anomaly detection results, quantization parameters, risk levels, and recommended actions.

[0086] The report generation module integrates the reasoning results from multiple task branches to generate visual profile images with risk level rendering, statistical tables of abnormal results that conform to industry standards, and standardized detection and diagnostic reports.

[0087] The aforementioned system, through the coordinated operation of the data acquisition module, the model inference module with a built-in multi-task learning model, and the report generation module, achieves fully automated processing from raw detection data to standardized detection and diagnostic reports. It compresses the time of 2-4 hours for traditional manual interpretation of 100-meter-level profiles to the second level, while outputting standardized results that meet the requirements of Appendix G of JGJ / T 437-2018, directly supporting engineering decision-making and closed-loop management.

[0088] This embodiment also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements all the steps of the ground-penetrating radar underground anomaly collaborative risk assessment method. The memory is used to store the computer program and model weight files. The processor can be any one or a combination of a central processing unit (CPU), a graphics processing unit (GPU), a neural network processor (NPU), or a field-programmable gate array (FPGA).

[0089] This embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements all the steps of the ground-penetrating radar underground anomaly collaborative risk assessment method. The computer-readable storage medium can be a non-volatile storage medium, including but not limited to hard disks, solid-state drives, USB flash drives, SD cards, optical discs, etc.

[0090] Example 1: Application of Underground Defect Detection and Risk Assessment in Urban Main Roads

[0091] Step 1: Data Acquisition and Preprocessing

[0092] The Impulse RADAR 4080 ground-penetrating radar, equipped with 400MHz and 800MHz shielded antennas, was used to continuously collect data along the entire length of the city's main roads. The data acquisition parameters were set as follows: acquisition step size 2cm, time window 150ns. A total of 1200 effective profiles were collected, covering typical underground defects such as cavities, voids, loose bodies, and water-rich bodies.

[0093] Preprocessing of the raw data includes background removal, 50-800MHz bandpass filtering, automatic gain compensation, zero-point correction, and data normalization. Calibration is performed using a DN300 water supply pipeline at a known burial depth, yielding an equivalent electromagnetic wave velocity v = 0.105 m / ns in the detection area. A time-depth conversion mapping is established, enabling the subsequent quantization branch to convert pixel dimensions in image space to geometric dimensions in physical space.

[0094] Step 2: Dataset Construction and Model Training

[0095] 1000 preprocessed profile images were selected from 1200 profile groups, and four layers of annotation were completed by 3-4 senior geophysical engineers and risk assessment experts. The four layers of annotation specifically include: anomaly detection bounding boxes and category labels, pixel-level segmentation mask labels, geometric quantization parameter ground truth labels, and risk value and level ground truth labels calculated according to industry standards. The annotated samples were divided into a training set (700 images), a validation set (200 images), and a test set (100 images) in a 7:2:1 ratio.

[0096] The specific configuration of the model is as follows: The shared backbone network adopts EfficientNet-B4, which is fine-tuned by transfer learning based on ImageNet pre-trained weights; the anomaly detection and classification branch adopts the YOLOv8 detection head, which integrates channel attention and spatial attention modules; the pixel-level segmentation and quantization branch adopts the DeepLabv3+ encoder-decoder structure, which shares the bottom to middle layer features with the detection branch to achieve mutual supervision learning; the risk assessment branch adopts a 3-layer multilayer perceptron (MLP) lightweight network, whose computational logic embeds the risk assessment system of "Technical Specification for Detection of Urban Underground Diseases" DB11 / T 1399-2017 and "Technical Standard for Comprehensive Detection and Risk Assessment of Urban Underground Diseases" JGJ / T 437-2018.

[0097] Training parameter settings: Using the AdamW optimizer, initial learning rate 1e-4, batch size 8, and 100 training epochs. An adaptive weighted total loss function is used. .

[0098] in, For CIoU loss, For Dice's loss, For cross-entropy loss, , , The initial values ​​are 0.4, 0.4, and 0.2, respectively, and are dynamically adjusted during training based on the convergence of each branch.

[0099] The evaluation results on the test set are as follows: anomaly classification accuracy of 94.5%, average intersection-union ratio (mIoU) of segmentation of 89.8%, average absolute error (MAE) of the top burial depth of the anomaly target of 4.1 cm, the consistency between the risk rating results and the manual assessment results of senior experts of 92.9%, and the average inference time of a single GPR image of 0.75 seconds.

[0100] Step 3: On-site Engineering Application

[0101] like Figure 5As shown, before the completion and acceptance of a newly built main road, a trained inference model was used to automatically interpret the GPR data of the entire line in batches. Taking a typical profile as an example, the original GPR images, after preprocessing, exhibited two types of typical anomalous signals: hyperbolic reflection signals representing point-like anomalies, and layered reflection signals caused by the distortion or discontinuity of the phase axis due to changes in the medium. The preprocessed GPR images and environmental context information were input into the inference model, and all inference was completed synchronously through a single forward propagation.

[0102] Anomaly detection and classification branches identified and located two anomalous targets: one was a cavity (confidence 0.95), and the other was a loose body (confidence 0.88). Pixel-level segmentation and quantization branches simultaneously generated pixel-level segmentation masks for the two targets, and combined with temporal-depth transformation mapping to calculate geometric quantization parameters: the cavity's top is buried at a depth of approximately 1.2m, with a projected area of ​​approximately 2.5m², located directly below the roadway; the loose body's top is buried at a depth of approximately 2.1m, with a projected area of ​​approximately 4.3m².

[0103] The risk assessment branch comprehensively considers geometric quantitative parameters, category labels, and environmental context information (road load level is urban arterial road, surrounding area is densely commercial area), and assesses the risk level according to embedded industry standards: voids are rated as Level V (extremely high risk), highlighted in red, and it is recommended to immediately cordon off and carry out engineering treatment; loose bodies are rated as Level III (relatively high risk), highlighted in yellow, and it is recommended to include them in regular tracking and monitoring.

[0104] The system ultimately generates a visual image with risk level rendering, an abnormal result statistical table that meets the requirements of Appendix G of JGJ / T 437-2018, and a standardized test and diagnosis report, which are then directly delivered to the owner for use.

[0105] Step 4: On-site verification

[0106] Drilling was conducted to verify the cavity anomalies identified and assessed as Level V risk by the model. The drilling results showed that the actual cavity depth was approximately 1.3m, which was basically consistent with the model's prediction of a top depth of 1.2m. The deviation was within an acceptable range, confirming the effectiveness and high accuracy of the method of this invention in practical engineering scenarios.

[0107] Example 2: Application of Grouting Quality Assessment Behind Shield Tunnels in Soft Soil Areas

[0108] Step 1: Data Acquisition and Preprocessing

[0109] An Impulse RADAR 1760 ground-penetrating radar, equipped with 170MHz and 600MHz shielded antennas, was used to collect grouting detection profiles along the tunnel circumference. Calibration was performed using 35cm thick concrete segments, yielding an equivalent electromagnetic wave velocity v=0.09m / ns for the soft soil layer. Based on this, a time-depth conversion mapping relationship under soft soil layer conditions was established.

[0110] Step 2: Model Fine-tuning and Inference

[0111] Based on the inference model trained above, 200 sets of labeled data for tunnel wall grouting detection scenarios were added for transfer fine-tuning. Fine-tuning can be performed using a relatively small learning rate (1×10). -5 The model was trained for 30 epochs to maintain its existing capabilities in road scenarios while adapting to the data distribution characteristics of tunnel scenarios. After fine-tuning, batch inference was performed on the detection profiles of all tunnels along the entire line.

[0112] Step 3: Application Results

[0113] Application results show that the model achieves a 93.6% accuracy rate in identifying grouting defects behind tunnel walls, with a quantification error of grout thickness ≤5cm, and a 91.3% consistency between the risk rating results and the manual assessment results from the tunnel engineering expert group. Through the automated interpretation and risk assessment of this invention, the engineering pain points of "invisible, difficult to detect, and difficult to classify" grouting quality behind shield tunnel walls in soft soil areas are solved, realizing full-process automation of grouting quality assessment.

[0114] The above embodiments fully demonstrate that the method of the present invention can efficiently, accurately, and in a standardized manner complete the intelligent interpretation and risk assessment of GPR detection data, and has strong engineering application value and large-scale promotion prospects.

[0115] The above description is merely a description of preferred embodiments of the present invention and is not intended to limit the scope of the present invention in any way. Any changes or modifications made by those skilled in the art based on the above disclosure shall fall within the protection scope of the claims.

Claims

1. A method for collaborative risk assessment of underground anomalies using ground-penetrating radar, characterized in that, Includes the following steps: S1: Acquire detection data and environmental context information of the detection area; S2: Construct a multi-task learning model, which includes a shared feature extraction network and multiple parallel and collaborative task branches. The multiple task branches include at least an anomaly detection and classification branch, a pixel-level segmentation and quantization branch, and a risk assessment branch. S3: Train the multi-task learning model using a multi-task labeled dataset and an adaptive weighted total loss function to obtain the inference model; S4: Input the detection data and environmental context information into the inference model, and through one forward propagation, the inference results are synchronously output by the multiple task branches; S5: Integrate the inference results from multiple task branches to generate a detection and diagnostic report.

2. The ground-penetrating radar underground anomaly collaborative risk assessment method as described in claim 1, characterized in that, In step S1: The detection data is ground-penetrating radar B-Scan data; and / or, After acquiring the detection data, the process further includes preprocessing the data and obtaining the equivalent electromagnetic wave velocity of the underground medium in the detection area through calibration, in order to establish a time-depth conversion mapping relationship between the radar two-way travel time and the actual burial depth; and / or, The environmental context information includes at least one of the following: load conditions, geological conditions, and surrounding environmental attributes of the area where the abnormal target is located.

3. The ground-penetrating radar underground anomaly collaborative risk assessment method as described in claim 2, characterized in that, The preprocessing includes at least one of the following: DC offset removal, zero-point correction, background noise removal, bandpass filtering, time gain compensation, and data normalization.

4. The ground-penetrating radar underground anomaly collaborative risk assessment method as described in claim 1, characterized in that, The shared feature extraction network in step S2 adopts a deep convolutional neural network or a visual Transformer network pre-trained on a general image dataset. It is fine-tuned on the special dataset of the probe data through transfer learning, and the extracted shared features are adapted to the feature requirements of the multiple task branches.

5. The ground-penetrating radar underground anomaly collaborative risk assessment method as described in claim 1, characterized in that, The multiple task branches in step S2 are specifically as follows: The anomaly detection and classification branch adopts an integrated attention mechanism target detection architecture to output the bounding boxes, category labels, and classification confidence scores of underground anomalies in the detection data. The pixel-level segmentation and quantization branch uses a semantic segmentation network with an encoder-decoder structure to output pixel-level segmentation masks for anomalous targets and automatically calculate their geometric quantization parameters. The risk assessment branch employs a lightweight multilayer perceptron network. Its inputs include at least the geometric quantization parameters, the category labels, and the environmental context information. Its outputs include the risk level and corresponding handling recommendations.

6. The ground-penetrating radar underground anomaly collaborative risk assessment method as described in claim 5, characterized in that, The object detection architecture with the integrated attention mechanism is any one of the improved YOLOv8, YOLOv10, or RetinaNet; and / or, The category labels include at least six categories: voids, cavities, loose bodies, water-rich bodies, pipelines, and reinforcing steel; and / or, The semantic segmentation network is DeepLabv3+ or U-Net++; and / or, The geometric quantification parameters include the top burial depth, bottom burial depth, vertical height, horizontal span, and projected area of ​​the anomalous target; and / or, The calculation logic of the risk assessment branch embeds the risk assessment system of the "Technical Specification for Detection of Urban Underground Diseases" DB11 / T 1399-2017 and / or the "Technical Standard for Comprehensive Detection and Risk Assessment of Urban Underground Diseases" JGJ / T 437-2018, automatically calculates the probability level of risk occurrence, the risk consequence level, and the comprehensive risk value, and classifies the risk level accordingly.

7. The ground-penetrating radar underground anomaly collaborative risk assessment method as described in claim 1, characterized in that, The multi-task labeled dataset in step S3 contains four corresponding labels for each sample: anomaly detection box and category label, pixel-level segmentation mask label, geometric quantization parameter ground value label, and risk probability level, risk consequence level, comprehensive risk value and risk level ground value label calculated according to industry standards. And / or, The formula for the adaptive weighted total loss function in step S3 is: ； in, CIoU loss for anomaly detection and classification branches, The Dice loss is used for pixel-level segmentation and quantization branches. The cross-entropy loss for the risk assessment branch; , , These are adaptive weights for the losses of the three tasks, which can be dynamically adjusted during training.

8. The ground-penetrating radar underground anomaly collaborative risk assessment method as described in claim 1, characterized in that, In step S5, a visual image with anomaly annotations and risk level rendering is also generated, as well as an anomaly result statistics table that conforms to industry standards. The format and fields of the abnormal results statistics table match the requirements of Appendix G of the "Technical Standard for Comprehensive Detection and Risk Assessment of Urban Underground Diseases" JGJ / T 437-2018. The abnormal results statistics table shall at least include the fields of abnormality number, type, location, center coordinates, projected area, depth of impact, risk level, and treatment suggestion.

9. A ground-penetrating radar-based collaborative risk assessment system for underground anomalies, characterized in that, include: The data acquisition module is used to acquire detection data and environmental context information of the detection area; The model inference module has a built-in multi-task learning model, which includes a shared feature extraction network and multiple parallel and collaborative task branches. The multiple task branches include at least an anomaly detection and classification branch, a pixel-level segmentation and quantization branch, and a risk assessment branch. The model inference module is used to input the detection data and environmental context information into the multi-task learning model, and through one forward propagation, the multiple task branches synchronously output the inference results. The report generation module is used to integrate the reasoning results of multiple task branches and generate a detection and diagnostic report.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the ground-penetrating radar underground anomaly collaborative risk assessment method as described in any one of claims 1 to 8.

Images