Ground penetrating radar tunnel geology advanced forecasting method and system based on incremental learning

Abstract

The invention discloses a ground penetrating radar tunnel geology advanced forecasting method and system based on incremental learning, and aims to solve the problems that in the prior art, dynamic newly-added data cannot be utilized, a large number of labels are relied on, and disastrous forgetting exists. According to the method, an incremental learning mechanism and a target detection model are combined, and a course pseudo-label strategy is introduced, so that the model can efficiently learn from a GPR measured data stream which is continuously generated and contains a large amount of unlabeled data, the prediction precision and generalization ability are continuously improved, meanwhile, the forgetting of old knowledge is effectively relieved, and the prediction efficiency is improved under the support of a small amount of labels. According to the tunnel geological hazard intelligent identification and prediction method, continuous self-optimization can be carried out by utilizing dynamically added actually-measured data including a large amount of unlabeled data, memory of existing knowledge is kept while the method adapts to a new environment and a new hazard type, and finally high-precision, high-efficiency and high-generalization-ability tunnel geological hazard intelligent identification and prediction are achieved.

Description

[0001] This application relates to the field of safety and environmental protection construction testing technology, specifically to a ground-penetrating radar method and system for advanced geological prediction of tunnels based on incremental learning. Background Technology

[0002] In recent years, my country has ranked first in the world in both the number and total mileage of completed tunnels, and this trend is expected to continue. However, the risks of geological disasters such as landslides, water inrushes, and cavities behind the lining are becoming increasingly prominent during tunnel construction and operation, causing significant personnel and economic losses. Therefore, accurate and efficient geological forecasting during tunnel construction is a core element in disaster prevention and safety assurance. Ground Penetrating Radar (GPR), as a non-destructive geophysical exploration method, effectively detects adverse geological bodies such as rock structures, aquifers, karst caves, and fracture zones ahead of the tunnel face by emitting high-frequency electromagnetic waves and receiving reflected signals from the underground medium. It has become one of the mainstream technologies for tunnel geological forecasting. The resulting B-scan images are a direct reflection of the spatiotemporal response of electromagnetic waves. However, the interpretation of GPR data is highly dependent on the professional experience of interpreters, presenting inherent challenges such as strong subjectivity, low efficiency, and difficulty in identifying complex interference signals (i.e., the "multiple solutions" problem). Especially in tunnel environments with complex media conditions and varied target shapes, the accuracy and consistency of manual interpretation are difficult to guarantee. In recent years, machine learning, particularly deep learning, has made breakthroughs in image recognition and target detection, providing new avenues for intelligent interpretation of GPR data. Some studies have attempted to apply models such as convolutional neural networks, Faster R-CNN, and YOLO to the automatic identification of hazardous targets in GPR images, achieving performance superior to traditional methods. However, existing deep learning-based GPR interpretation schemes typically suffer from the following key drawbacks:

[0003] First, there is the problem of static model training and insufficient data utilization: most existing methods adopt a "one-off" training model. That is, before model deployment, a fixed and limited dataset (usually containing a large amount of numerical simulation data and a small amount of measured data) is used for training and validation, and the model parameters are then frozen. Tunnel engineering is a typical dynamic and continuous process. As tunneling progresses, a large amount of new measured GPR data reflecting different surrounding rock conditions is continuously generated. This dynamically added data contains rich field information, but because traditional models lack continuous learning capabilities, they cannot use this new data to optimize and update themselves, resulting in limited model generalization ability and difficulty in adapting to the forecasting needs under long-distance and variable geological conditions.

[0004] Secondly, there is a bottleneck in data annotation: Supervised learning is the current mainstream paradigm, and its performance heavily relies on large-scale, high-quality annotated data. However, accurate geological hazard annotation of GPR images requires geophysical experts to spend a significant amount of time and effort, resulting in high costs. In engineering practice, newly acquired data often cannot be annotated in a timely manner, resulting in a large amount of "unannotated data." Traditional methods either discard the data, leading to waste, or require waiting for manual annotation to be completed before retraining the model, a lengthy process that cannot meet the timeliness requirements of construction.

[0005] Finally, there's the catastrophic forgetting problem: simply mixing old and new data and retraining the model, while enabling it to learn new knowledge, often severely overwrites or forgets previously learned knowledge, leading to a sharp decline in performance on previously seen data. This is unacceptable in dynamic, incremental engineering scenarios. Incremental learning, also known as continuous learning, aims to enable models to continuously learn new knowledge from a continuous stream of data, much like humans, while retaining as much of the previously learned knowledge as possible. This provides a theoretical framework for solving the dynamic adaptability problem in GPR intelligent interpretation. However, directly applying existing incremental learning algorithms to GPR tunnel geological prediction faces unique challenges:

[0006] 1) GPR images have a relatively low signal-to-noise ratio, and interference (such as reflection from lining steel bars, system noise, etc.) is mixed with real hazard signals;

[0007] 2) The reflectance characteristics of different geological hazards (such as cavities, water-bearing areas, and fracture zones) vary greatly, and the sample size may be naturally unbalanced.

[0008] 3) Under the constraint of scarce labels, it is necessary to make efficient use of unlabeled data.

[0009] Therefore, there is an urgent need for an innovative incremental learning intelligent interpretation scheme specifically designed for ground-penetrating radar tunnel geological advance prediction scenarios. Summary of the Invention

[0010] The main objective of this application is to provide a ground-penetrating radar (GPR) method for advanced geological prediction of tunnels based on incremental learning, comprising the following steps:

[0011] S1: Obtain B-scan image data of tunnel geology collected by ground penetrating radar, and construct an initial training dataset with geological hazard labels;

[0012] S2: Based on the initial training dataset, supervised training is performed on the object detection basic model built on the YOLOv5 architecture to obtain the trained initial model;

[0013] S3: After the initial model is deployed, acquire the newly added B-scan image data, wherein the newly added B-scan image data includes unlabeled data and partially labeled data;

[0014] S4: Employ a course pseudo-labeling strategy to process the unlabeled data and dynamically generate pseudo-labels;

[0015] S5: Construct a hybrid training set using the newly added labeled data and the generated pseudo-labeled data;

[0016] S6: Starting with the weights of the initial model, the model is incrementally trained based on the mixed training set. The model parameters are updated by optimizing the total loss function that combines supervised loss and unsupervised loss to obtain the incrementally optimized model. The supervised loss is calculated based on labeled data, and the unsupervised loss is calculated based on pseudo-labeled data.

[0017] S7: Use the incremental optimization model to detect and identify geological hazards in the new ground-penetrating radar B-scan images, and output hazard category and location information.

[0018] In one embodiment, the target detection model built on the YOLOv5 architecture includes the following network structure:

[0019] The input layer is used to perform Mosaic data augmentation, adaptive anchor box calculation, and adaptive image scaling on the input B-scan image.

[0020] The backbone network, using the CSP-Darknet53 structure, is used to extract multi-scale features from the input image;

[0021] The neck network employs a CSP-PAN structure that includes an SPPF module to fuse and enhance feature information from different scales of the backbone network.

[0022] The head network is used for target bounding box regression, confidence prediction, and category classification based on the fused features.

[0023] In one embodiment, the preprocessing of the B-scan image data further includes one or more data enhancement operations such as random flipping, random cropping, and color space adjustment.

[0024] In one embodiment, the step of using a course pseudo-labeling strategy to process the newly added unlabeled data and dynamically generate pseudo-labels specifically includes:

[0025] The newly added unlabeled data is input into the current model for prediction, and the prediction confidence scores for each category are obtained.

[0026] Based on a preset fixed threshold τ, the number of samples with prediction confidence exceeding τ in each category c is counted as σ_t(c), which is used as an estimate of the learning effect of that category.

[0027] The learning effect estimates σ_t(c) are normalized to obtain the normalization coefficient β_t(c);

[0028] Based on the normalization coefficient β_t (c) and the fixed threshold τ, the dynamic threshold T_t (c) = β_t (c) *τ for each category c in the current learning stage is dynamically calculated.

[0029] Prediction results with a confidence level higher than the corresponding class dynamic threshold T_t (c) are used as pseudo-labels for the sample.

[0030] In one embodiment, the total loss function is a weighted sum of supervised loss and unsupervised loss, expressed as:

[0031] Total Loss = A * Loss_box + B * Loss_obj + C * Loss_cls +λ* Loss_unsup;

[0032] Wherein, Loss_box is the localization loss, which adopts CIoU Loss; Loss_obj is the confidence loss, and Loss_cls is the classification loss, both of which adopt binary cross-entropy loss; Loss_unsup is the unsupervised loss calculated based on pseudo-labels; A, B, and C are the weight coefficients of localization loss, confidence loss, and classification loss, respectively, and λ is the balancing weight of unsupervised loss.

[0033] In one embodiment, the formula for calculating CIoU Loss is:

[0034] CIoU Loss =

[0035] Where IoU is the intersection-union ratio between the predicted bounding box and the ground truth bounding box. Let be the Euclidean distance between the centers of the predicted bounding box and the ground truth bounding box, and let c be the length of the diagonal of the smallest closed region enclosing the predicted bounding box and the ground truth bounding box. These are the weighting coefficients. This is a parameter used to measure the consistency of the aspect ratio between the predicted bounding box and the ground truth bounding box.

[0036] In one embodiment, the method further includes:

[0037] The performance of the incremental optimization model is evaluated based on the confusion matrix, F1-confidence curve, and detection errors of the old and new tasks. The balance weights λ of the fixed threshold τ or unsupervised loss are then adjusted according to the evaluation results.

[0038] A ground-penetrating radar tunnel geological advance prediction system based on incremental learning, the system comprising:

[0039] The data acquisition and preprocessing module is used to acquire and preprocess tunnel B-scan image data collected by ground penetrating radar, and to build the initial training dataset and dynamically added datasets.

[0040] The initial model training module is used to perform supervised training on the YOLOv5-based object detection base model based on an initial training dataset with labeled parameters, and generate the initial model.

[0041] The course pseudo-label generation module is used in the incremental learning phase to dynamically calculate the dynamic thresholds of each category based on the current model's prediction results for newly added unlabeled data, and then filter and generate pseudo-labels.

[0042] The incremental learning optimization module is used to initialize the model with the parameters of the initial model, combine the newly added labeled data and the generated pseudo-label data, and perform incremental training on the model by optimizing the total loss function to update the model parameters.

[0043] The intelligent geological hazard interpretation module is used to analyze and interpret the input ground-penetrating radar B-scan images in real time using a trained incremental optimization model, and automatically identify and locate geological hazard targets within them.

[0044] In one embodiment, the course pseudo-tag generation module specifically includes:

[0045] The learning effect estimation unit is used to count the number of effective prediction samples σ_t (c) for each category based on the current model prediction results and a fixed threshold τ.

[0046] The normalization processing unit is used to normalize the learning effect estimates σ_t(c) for each category to obtain β_t(c);

[0047] The dynamic threshold calculation unit is used to calculate the dynamic threshold for each category according to the formula T_t (c) = β_t (c) * τ;

[0048] The pseudo-label filtering unit is used to identify prediction results with a prediction confidence higher than the corresponding category dynamic threshold T_t (c) as pseudo-labels.

[0049] In one embodiment, the total loss function in the incremental learning optimization module is a weighted sum of CIoU localization loss based on labeled data, binary cross-entropy confidence loss and classification loss, and unsupervised loss based on pseudo-label data.

[0050] Therefore, this application has the following beneficial effects:

[0051] This application provides a ground-penetrating radar (GPR) method for advanced geological prediction of tunnels based on incremental learning, comprising the following steps: acquiring B-scan image data of tunnel geology collected by GPR, and constructing an initial training dataset with geological hazard labels; performing supervised training on a target detection basic model built on the YOLOv5 architecture based on the initial training dataset to obtain a trained initial model; after deploying the initial model, acquiring newly added B-scan image data, wherein the newly added B-scan image data includes unlabeled data and partially labeled data; using a course pseudo-label strategy to process the unlabeled data and dynamically generate pseudo-labels; constructing a hybrid training set using the newly added labeled data and the generated pseudo-labeled data; starting from the weights of the initial model, incrementally training the model based on the hybrid training set, updating the model parameters by optimizing the total loss function that combines supervised loss and unsupervised loss, and obtaining an incrementally optimized model; wherein the supervised loss is calculated based on the labeled data, and the unsupervised loss is calculated based on the pseudo-labeled data; using the incrementally optimized model to detect and identify geological hazards in the new GPR B-scan images, and outputting hazard category and location information. This application discloses a ground-penetrating radar (GPR) method and system for advanced geological prediction of tunnels based on incremental learning, aiming to solve the problems of existing technologies, such as inability to utilize dynamically added data, reliance on large amounts of labeled data, and catastrophic forgetting. This application combines an incremental learning mechanism with a target detection model, introducing a course pseudo-label strategy to enable the model to learn efficiently from continuously generated GPR measured data streams containing a large amount of unlabeled data. This continuously improves prediction accuracy and generalization ability, while effectively mitigating the forgetting of old knowledge. With limited labeling support, it can continuously self-optimize using dynamically added measured data, including a large amount of unlabeled data, adapting to new environments and new hazard types while maintaining the retention of existing knowledge. Ultimately, this achieves high-precision, high-efficiency, and high-generalization intelligent identification and prediction of tunnel geological hazards. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is a system flowchart of a ground-penetrating radar tunnel geological advance prediction method based on incremental learning;

[0054] Figure 2This is a flowchart of the incremental learning training process for a ground-penetrating radar tunnel geological advance prediction method based on incremental learning.

[0055] Figure 3 This is the confusion matrix evaluation curve of the ground-penetrating radar tunnel geological advance prediction method based on incremental learning;

[0056] Figure 4 This is a numerical simulation example verification diagram of the ground-penetrating radar tunnel geological advance prediction method based on incremental learning;

[0057] Figure 5 This is a comparison of model predictions from a ground-penetrating radar tunnel geological advance prediction method based on incremental learning. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0059] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of this application.

[0060] To address the shortcomings of existing technologies, this application provides a ground-penetrating radar (GPR) method for advanced geological prediction of tunnels based on incremental learning. It trains an initial target detection model based on YOLOv5 using initially labeled data. The key innovation lies in the subsequent incremental learning mechanism: facing the continuous influx of new data (mostly unlabeled) during construction, the system introduces a course pseudo-label strategy to dynamically evaluate the model's learning status for each disease category. Based on this, it adaptively sets differentiated confidence thresholds for each category, thereby efficiently and automatically generating training pseudo-labels from massive amounts of unlabeled data. Subsequently, starting with historical model parameters, and combining a small amount of newly labeled data with a large amount of generated pseudo-label data, the model is incrementally trained and optimized by optimizing a composite objective function that integrates supervised and unsupervised losses. This process enables the model to absorb new knowledge (adapting to new geological conditions and new disease characteristics) while retaining old knowledge to the greatest extent possible, effectively mitigating catastrophic forgetting. This application realizes a paradigm shift in ground-penetrating radar interpretation models from "one-time training and static use" to "continuous learning and dynamic evolution," significantly reducing the reliance on manual annotation and improving the model's generalization ability and long-term forecast accuracy under long tunnels and complex geological conditions.

[0061] This application provides a ground-penetrating radar tunnel geological advance prediction method based on incremental learning, including steps S1 to S7, referring to... Figure 1 , Figure 1 This is a system flowchart of a ground-penetrating radar tunnel geological advance prediction method based on incremental learning.

[0062] S1: Obtain B-scan image data of tunnel geology collected by ground penetrating radar, and construct an initial training dataset with geological hazard labels;

[0063] S2: Based on the initial training dataset, supervised training is performed on the object detection basic model built on the YOLOv5 architecture to obtain the trained initial model;

[0064] S3: After the initial model is deployed, acquire the newly added B-scan image data, wherein the newly added B-scan image data includes unlabeled data and partially labeled data;

[0065] S4: Employ a course pseudo-labeling strategy to process the unlabeled data and dynamically generate pseudo-labels;

[0066] S5: Construct a hybrid training set using the newly added labeled data and the generated pseudo-labeled data;

[0067] S6: Starting with the weights of the initial model, the model is incrementally trained based on the mixed training set. The model parameters are updated by optimizing the total loss function that combines supervised loss and unsupervised loss to obtain the incrementally optimized model. The supervised loss is calculated based on labeled data, and the unsupervised loss is calculated based on pseudo-labeled data.

[0068] S7: Use the incremental optimization model to detect and identify geological hazards in the new ground-penetrating radar B-scan images, and output hazard category and location information.

[0069] Specifically, in this embodiment, this application provides a ground-penetrating radar (GPR) method for advanced geological prediction of tunnels based on incremental learning. This method aims to address the technical challenges of existing static models being unable to utilize continuously generated data during construction, over-reliance on manual annotation, and the existence of catastrophic forgetting. By integrating an incremental learning framework with a course pseudo-labeling strategy, this method enables the intelligent interpretation system to dynamically evolve, continuously improving prediction accuracy and adaptability. The following, in conjunction with the appendix... Figure 1 The system flowchart shown illustrates steps S1 to S7 in detail.

[0070] S1: Obtain B-scan image data of tunnel geology collected by ground penetrating radar, and construct an initial training dataset with geological hazard labels.

[0071] Ground-penetrating radar (GPR) equipment is used to scan the tunnel area and collect raw electromagnetic wave signals. These signals are then preprocessed (e.g., gain adjustment, background denoising, filtering, and time-depth conversion) to convert them into two-dimensional B-scan images, which visually reflect the distribution of electromagnetic properties of the underground medium. B-scan is a two-dimensional display method that forms a cross-sectional image along a certain direction through the spatial arrangement of multiple A-scan data. Its core principle is to arrange and combine multiple A-scans (amplitude scans, one-dimensional signals) according to their spatial positions to form a two-dimensional image. The initial dataset comes from historical engineering archives or the early survey stage. Geological and geophysical experts manually interpret and label the geological hazards (such as cavities, water-rich areas, fractured rock zones, weak interlayers, etc.) based on image features. The labeling information includes the hazard category label and its precise bounding box coordinates in the image. To ensure the initial robustness of the model, the initial dataset needs to cover a certain degree of scene diversity, such as image samples under different surrounding rock grades and different water content conditions. These labeled image-label pairs form the basis of supervised learning and are used to train the model's initial recognition ability.

[0072] S2: Based on the initial training dataset, supervised training is performed on the object detection basic model built on the YOLOv5 architecture to obtain the trained initial model.

[0073] The YOLOv5 object detection architecture was chosen due to its good balance between accuracy and speed, making it suitable for on-site deployment in engineering projects. First, the initial training dataset was divided into training and validation sets. Before training, Mosaic data augmentation was used, randomly stitching together four training images to simulate complex background combinations and improve the model's robustness to small targets and background interference. Simultaneously, adaptive anchor box calculation was performed, automatically optimizing the size of preset anchor boxes based on the size distribution of real bounding boxes in the dataset to better match the typical scale of geological hazard targets in tunnels. During training, the backbone network CSP-Darknet53 efficiently extracts multi-level features from B-scan images; the neck network CSP-PAN, combined with the SPPF module, fuses and enhances deep and shallow features, enabling the model to detect hazard targets at different scales; the head network outputs the target's class probability, confidence score, and bounding box coordinates. The CIoU loss function was used to optimize bounding box regression, which comprehensively considers overlap area, center point distance, and aspect ratio for more accurate localization. Binary cross-entropy was used for classification and confidence loss. Iterative training is performed by setting appropriate learning rate strategies (such as Cosine decay) and optimizers until the model's performance metrics (such as mean average precision, mAP) on the validation set stabilize. The model obtained after training convergence is the "initial model" with basic prediction capabilities.

[0074] S3: After the initial model is deployed, acquire the newly added B-scan image data, wherein the newly added B-scan image data includes unlabeled data and partially labeled data.

[0075] The initial model trained in step S2 is integrated into the intelligent interpretation system at the tunnel construction site. As the tunnel continues to be excavated, the ground-penetrating radar performs a new round of detection according to a preset cycle (e.g., every cycle of excavation), continuously generating new B-scan image data. Due to the time-consuming and costly nature of expert annotation, only a small portion of this new data can be annotated in a timely and accurate manner (e.g., prioritizing the annotation of images of key sections or suspected high-risk areas), forming a small amount of "newly annotated data." The vast majority of newly acquired images are in an "unannotated" state in the initial stage. The system needs to establish a data management mechanism that can distinguish and store these two types of new data to provide raw materials for subsequent incremental learning. This step reflects the targeted solution of this invention to the current situation of "data annotation lag" and "scarcity of annotation resources" in actual engineering.

[0076] S4: Employ a course pseudo-labeling strategy to process the unlabeled data and dynamically generate pseudo-labels.

[0077] This step aims to utilize massive amounts of unlabeled data efficiently and with high quality. The course pseudo-labeling strategy mimics the human learning process of "from easy to difficult," dynamically adjusting the confidence threshold for assigning appropriate "pseudo-labels" to unlabeled data.

[0078] Prediction and Statistics: The newly added unlabeled data from the current batch are input into the current optimal model (the initial model obtained in S2 in the initial stage) for inference, obtaining the prediction confidence of each sample for each hazard category. For a preset global fixed threshold τ (e.g., 0.7), the number of samples with prediction confidence exceeding τ in each category c is counted as σ_t(c). The higher the value of σ_t(c), the more confident the model is in identifying that category, i.e., the better its "learning effect".

[0079] Dynamic threshold calculation: To apply different screening strictness to categories at different learning states, σ_t(c) is normalized: β_t(c) = σ_t(c) / max(σ_t). Then, a fixed threshold τ is scaled using β_t(c) to obtain the dynamic threshold T_t(c) = β_t(c) * τ for each category. For categories with good learning performance, β_t(c) is close to 1, and T_t(c) is close to τ, indicating strict screening and ensuring high reliability of pseudo-labels. For categories with poor learning performance (difficult to identify), β_t(c) is smaller, and T_t(c) is reduced, allowing more samples belonging to that category with slightly lower confidence to participate in training, thereby promoting the model to learn the difficult features of that category.

[0080] Pseudo-label generation: Iterate through all unlabeled samples. For a sample, if it is predicted to be of class c and the prediction confidence is higher than the dynamic threshold T_t(c) of that class, then accept the prediction and use its class and bounding box as the "pseudo-label" for that sample. Otherwise, the sample is not used in this round of training. This strategy effectively avoids noise introduced by low-quality pseudo-labels and ensures the balance and asymptotic nature of the training data distribution.

[0081] S5: Construct a hybrid training set using the newly added labeled data and the generated pseudo-labeled data.

[0082] The newly labeled data (high value, small quantity) obtained in step S3 is combined with the pseudo-labeled data (low cost, large quantity) generated in step S4 to form a "hybrid training set" for this round of model updates. This hybrid training set contains both authoritative supervision signals from expert knowledge and potential patterns mined by the model itself from a large amount of new data, providing a data foundation for the model to safely and effectively expand its knowledge.

[0083] S6: Starting with the weights of the initial model, incrementally train the model based on the mixed training set. By optimizing the total loss function that combines supervised and unsupervised loss, update the model parameters to obtain the incrementally optimized model.

[0084] Unlike traditional retraining, incremental training uses the already deployed initial model weights as the starting point for optimization, rather than random initialization. This is equivalent to fine-tuning and expanding upon the existing knowledge structure. Training is performed using the hybrid training set constructed in step S5, with the total loss function L_total designed as a weighted sum of the supervised loss L_sup and the unsupervised loss L_unsup: L_total = L_sup + λ * L_unsup.

[0085] Supervision loss L_sup: For real labeled data in the mixed training set, the calculation method is the same as in step S2, including classification loss, CIoU localization loss and confidence loss.

[0086] Unsupervised loss L_unsup: For pseudo-labeled data in the mixed training set, calculate the difference between the model's predicted value and the pseudo-label. Its calculation form can be the same as L_sup, but it should be noted that during backpropagation, the pseudo-label should be treated as a fixed target and its gradient should not be calculated.

[0087] The coefficient λ is used to adjust the contribution of unsupervised loss and prevent low-quality pseudo-labels from dominating the training direction in the early stages.

[0088] By optimizing L_total through backpropagation, the model consolidates existing knowledge (reflected by the initial weights and the original knowledge structure) while updating parameters using newly added real annotations and high-quality pseudo-annotations. This training method learns new scenarios and features while minimizing the "catastrophic forgetting" of old knowledge, achieving stable and efficient incremental learning. After training, a "superiorly optimized model" with improved performance and greater adaptability is obtained.

[0089] S7: Use the incremental optimization model to detect and identify geological hazards in the new ground-penetrating radar B-scan images, and output hazard category and location information.

[0090] The incremental optimization model obtained in step S6 is deployed to the front end of the forecasting system. For fresh B-scan images acquired in real time during subsequent tunnel excavation, the system can automatically and quickly perform inference analysis. Figure 2 This is a flowchart of the incremental learning training process for a ground-penetrating radar (GPR) method for advanced geological prediction of tunnels. The model outputs the category (e.g., "cavity"), confidence level, and pixel-level location (bounding box) of all detected geological hazards in the image. The system can visualize and overlay the results onto the original image, generating a forecast report containing information such as hazard type, approximate size, and distance from the tunnel face, providing timely, objective, and intelligent technical support for construction decisions. As the project progresses, the system can iteratively execute steps S3 to S7, enabling the model to continuously iterate and evolve, forming a closed-loop intelligent geological prediction system with lifelong learning capabilities.

[0091] In one embodiment, the target detection model built on the YOLOv5 architecture includes the following network structure:

[0092] The input layer is used to perform Mosaic data augmentation, adaptive anchor box calculation, and adaptive image scaling on the input B-scan image.

[0093] The backbone network, using the CSP-Darknet53 structure, is used to extract multi-scale features from the input image;

[0094] The neck network employs a CSP-PAN structure that includes an SPPF module to fuse and enhance feature information from different scales of the backbone network.

[0095] The head network is used for target bounding box regression, confidence prediction, and category classification based on the fused features.

[0096] Specifically, in this embodiment, the target detection model built on the YOLOv5 architecture has a specially adapted network structure to efficiently handle the unique characteristics of GPR B-scan images (such as complex texture features, variable target scale, and relatively low signal-to-noise ratio). This structure mainly consists of four collaboratively working parts: an input layer, a backbone network, a neck network, and a head network. Its core lies in transforming the original electromagnetic wave image into accurate geological hazard target detection results through a series of optimization designs.

[0097] The input layer is the first step in the model's processing of raw data. It is responsible for performing a series of preprocessing and enhancement operations on the input GPR B-scan images to improve the efficiency and robustness of subsequent learning. Specifically, this includes:

[0098] Mosaic data augmentation: This technique randomly scales, crops, and stitches four training images into a single composite image. For tunnel GPR data, this effectively simulates complex and varied underground structural combinations (e.g., stitching together images containing cavities, water-bearing areas, and intact surrounding rock), greatly enriching the background and contextual information of the training samples. This forces the model to learn to locate and identify targets in more complex and realistically oriented environments, significantly enhancing the model's ability to detect small and partially occluded targets, and also acting as a regularization mechanism to reduce the risk of overfitting.

[0099] Adaptive Anchor Box Calculation: Anchor boxes are a set of pre-defined reference boxes used to predict target boundaries. Traditional methods use fixed anchor box sizes, which may not match the actual size distribution of targets in a specific dataset. At the start of training, YOLOv5 runs a clustering algorithm (such as K-means) on the initial training dataset, automatically analyzing the aspect ratio of all ground truth labeled bounding boxes. This results in nine sets of prior anchor box sizes that best match the sizes of typical geological hazard targets (such as striped fracture zones and near-circular cavities) in the current GPR data. This adaptive process eliminates the tedious manual design, allowing the model to perform bounding box regression more efficiently from the beginning of training, accelerating convergence.

[0100] Adaptive Image Scaling: During model inference, the input image sizes may vary. Traditional methods involve padding or scaling the image to a uniform size, but this can introduce a large number of invalid gray borders, increasing computational burden. Adaptive scaling strategies aim to scale the image to a standard input size (e.g., 640x640) while minimizing the added padding and black borders. For GPR B-scan images, this helps preserve effective information about temporal depth relationships in the original profile and reduces unnecessary computational overhead, which is beneficial for subsequent real-time on-site processing.

[0101] The backbone network employs an improved CSP-Darknet53 as its feature extractor. This network is a deep convolutional neural network, its core being a combination of the CSPNet (Cross Stage Partial Network) structure and Darknet53. The CSP structure splits the feature map along the channel dimension, performing deep convolutional processing on one part and directly merging the other part with the deep processing result after a simple transformation. This design offers two major advantages: firstly, it significantly reduces computational complexity and memory consumption (approximately reducing the number of parameters by 30%), which is crucial for deployment on embedded devices at tunnel construction sites; secondly, it effectively mitigates the vanishing gradient problem and promotes feature reuse, enabling the network to extract rich multi-scale features from GPR images more deeply and efficiently. These features include both shallow details and textures (such as the phase changes of reflected waves at lesion boundaries) and deep semantic information (such as the overall morphological patterns of different geological structures).

[0102] The neck network employs a CSP-PAN structure including an SPPF module. Its primary task is to fuse and enhance feature information from different layers of the backbone network. The feature maps output by the backbone network have different spatial resolutions (scales): shallow features have high resolution and rich detail but weak semantics; deep features have low resolution, strong semantics, but lost detail. The CSP-PAN structure constructs a powerful feature pyramid through a top-down path (upsampling deep, semantically strong features and fusing them with shallow features) and a bottom-up path (downsampling the fused shallow, detailed features and fusing them again with deep features). This allows the model to possess rich semantic and detailed information across feature maps at different scales, enabling it to effectively detect targets with vastly different sizes in GPR images, such as large-scale fracture zones and small isolated holes. The SPPF module (Spatial Pyramid Pooling - Fast, SPPF) is a fast version of spatial pyramid pooling. By concatenating multiple max pooling layers of different sizes (such as 5x5, 9x9, 13x13), it can greatly expand the receptive field of the feature map without significantly increasing the computational cost, helping the model to better understand the global contextual information of the target. This is especially important for determining the geological significance of anomalous reflectors in GPR images.

[0103] The head network is the prediction part of the network, responsible for the final object detection based on the fully fused multi-scale feature maps output by the neck network. The head network works in parallel on three feature maps of different scales, respectively responsible for detecting small, medium, and large-sized targets. For each pre-defined anchor box location, the head network simultaneously outputs three types of information:

[0104] Bounding box regression: Predicts the offset of the target bounding box relative to the preset anchor box (center point coordinates, width, and height), and optimizes it using the CIoU loss function to achieve high-precision positioning.

[0105] Confidence prediction: Predicts the probability that the bounding box contains a geological hazard target (objectivity score).

[0106] Category Classification: Predicts the probability that the target belongs to various predefined geological hazard categories (such as cavities, water-rich areas, and fracture zones). YOLOv5 uses binary cross-entropy loss for multi-label classification, allowing a target to be assigned multiple possible category labels to accommodate the ambiguity or complex characteristics of certain geological phenomena.

[0107] Through this end-to-end design, the model can complete the location and identification of all targets at once, achieving a balance between detection speed and accuracy, and fully meeting the dual requirements of real-time performance and accuracy for tunnel geological advance prediction.

[0108] In one embodiment, the preprocessing of the B-scan image data further includes one or more data enhancement operations such as random flipping, random cropping, and color space adjustment.

[0109] Specifically, in this embodiment, the preprocessing workflow, in addition to conventional format conversion and normalization, systematically introduces a series of data augmentation operations. These operations artificially expand the diversity and complexity of the training data by applying a series of controllable random geometric and color transformations to the original images during the training phase, without altering their geological semantics. This effectively simulates various imaging variations that may be encountered in actual tunnel exploration, enabling the model to learn more fundamental and stable feature representations. The augmentation operations implemented in this embodiment include, but are not limited to, one or more combinations of random flipping, random cropping, and color space adjustment.

[0110] Random flipping: This operation includes horizontal and / or vertical flipping. For GPR B-scan images, the horizontal axis typically represents the spatial distance along the survey line, and the vertical axis represents the two-way travel time (reflecting depth). Horizontal random flipping simulates the situation where ground-penetrating radar probes back and forth along the same survey line or the survey line direction is defined in opposite directions, which helps the model understand the spatial symmetry or orientation-independent characteristics of geological structures. Vertical random flipping is not common in practical physics because the depth direction is irreversible, but as a strong data augmentation technique, it forces the model not to rely excessively on texture patterns in specific directions (such as the tilt direction of the phase axis of a primary reflection wave from a specific interface), but to focus on the relative intensity and morphological structure of the reflection wave group, thereby enhancing the model's ability to identify the morphological features of anomalies, rather than their absolute spatial orientation.

[0111] Random cropping: This operation randomly selects a sub-region from the original input image as a new training sample. During implementation, it must be ensured that the random cropping box contains at least one complete geological target (this can be controlled by the intersection-union (IU) threshold with the labeled box). This operation offers several benefits: First, it forces the model to learn to recognize targets within a local field of view, improving the model's sensitivity to small and partially visible targets; second, by cropping at different locations, it increases the diversity of target positions in the image, reducing the model's reliance on remembering fixed target locations; finally, cropping is essentially a simplified form of multi-scale training because the cropped region's proportion relative to the original image is random, which helps the model adapt to inputs with different resolutions or scales of interest.

[0112] Color space adjustment: Although GPR B-scan images are typically grayscale images, their pixel intensity values ​​represent the reflected intensity of electromagnetic waves. In this embodiment, "color space adjustment" mainly refers to random linear or nonlinear transformations of image brightness and contrast. For example, randomly adjusting the image's gamma value or performing histogram equalization within a certain range. This operation simulates the overall change in signal strength caused by differences in the attenuation characteristics of underground media, variations in radar system gain settings, or changes in environmental noise levels. By exposing the model to images of the same geological structure mapped with different intensities, the model's robustness to absolute reflection intensity values ​​can be enhanced, allowing it to focus more on the relative changes in reflected waves, waveform characteristics, and spatial combination patterns—more discriminative essential features—thereby reducing the risk of misjudgment due to slight inconsistencies in acquisition parameters. In each iteration of the training process, each input image undergoes random combinations of these transformations with a certain probability, generating a near-infinite number of derived training samples. This strategy greatly enriches the distribution of training data and is a key and efficient engineering technique for improving the generalization performance and stability of deep learning models, especially in tunnel geological prediction scenarios with limited initial labeled data.

[0113] In one embodiment, the step of using a course pseudo-labeling strategy to process the newly added unlabeled data and dynamically generate pseudo-labels specifically includes:

[0114] The newly added unlabeled data is input into the current model for prediction, and the prediction confidence scores for each category are obtained.

[0115] Based on a preset fixed threshold τ, the number of samples with prediction confidence exceeding τ in each category c is counted as σ_t(c), which is used as an estimate of the learning effect of that category.

[0116] The learning effect estimates σ_t(c) are normalized to obtain the normalization coefficient β_t(c);

[0117] Based on the normalization coefficient β_t (c) and the fixed threshold τ, the dynamic threshold T_t (c) = β_t (c) *τ for each category c in the current learning stage is dynamically calculated.

[0118] Prediction results with a confidence level higher than the corresponding class dynamic threshold T_t (c) are used as pseudo-labels for the sample.

[0119] Specifically, in this embodiment, the Curriculum Pseudo Labeling (CPL) strategy is the core algorithm for achieving efficient semi-supervised incremental learning under conditions of scarce labeling. The core idea of ​​this method is to mimic the human learning process of "from easy to difficult," dynamically evaluating the model's current mastery of each target category and adaptively and differentially setting the confidence threshold for assigning pseudo-labels to unlabeled data. This intelligently selects high-quality training samples while simultaneously promoting the model's learning of difficult samples and rare categories. The specific implementation steps are detailed below:

[0120] The newly acquired, unlabeled B-scan image data batches are input into the currently trained optimal model (the initial model in the first incremental cycle) for forward inference. For each unlabeled image, the model outputs its predicted probability distribution for all predefined geological hazard categories (such as "cavities," "water-rich areas," and "fracture zones"), as well as the confidence score for each potential target bounding box. This step aims to leverage the existing model's cognitive capabilities to generate preliminary judgments with a measure of uncertainty regarding unknown data.

[0121] To quantify the model's mastery of different categories, a pre-defined, globally uniform fixed confidence threshold τ (e.g., τ = 0.7) is introduced. This threshold represents a high confidence level benchmark. Subsequently, the prediction results of all unlabeled samples in this batch are iterated. For each category c, the number of samples predicted by the model as belonging to that category with a prediction confidence exceeding the fixed threshold τ is counted, denoted as σ_t(c). The value of σ_t(c) intuitively reflects the model's ability to identify category c under "high standards": the larger the number, the more samples of that category the model can identify with high confidence, i.e., the better its current "learning effect" on that category, and the more robust its feature extraction; conversely, it indicates that the target category may be more difficult to learn or its features are not yet clear.

[0122] Since the absolute values ​​of σ_t(c) can vary greatly between different categories, direct use is inconvenient. Therefore, normalization is performed:

[0123] β_t(c) = σ_t(c) / max(σ_t)

[0124] Where max(σ_t) is the maximum value of σ_t(c) among all categories in this batch. After normalization, the category with the best learning performance has β_t(c)=1, and the β_t(c) of the other categories is between 0 and 1. This constitutes a coefficient that measures the relative learning difficulty of each category.

[0125] Next, this coefficient is used to dynamically calculate the pseudo-label generation threshold for each category:

[0126] T_t(c) = β_t(c) * τ

[0127] For categories with good learning performance (β_t(c) close to 1) (such as "holes"), the dynamic threshold T_t(c) is close to or even equal to the strictly fixed threshold τ. This means that only predictions that the model is very confident about will be adopted as pseudo-labels, ensuring high reliability for adding new training samples to this class and preventing the introduction of noise.

[0128] For categories with poor learning performance (smaller β_t(c)) (such as "fractured zones"), the dynamic threshold T_t(c) is reduced accordingly. This effectively relaxes the selection criteria, allowing more predictions belonging to this category but with relatively low model confidence (i.e., "hard samples" or "boundary samples") to generate pseudo-labels and be added to the training. This forces the model to be exposed to more challenging features related to this category in subsequent training, thereby specifically improving its learning ability for this weak category.

[0129] Finally, the prediction for each unlabeled sample is finalized based on the calculated category-specific dynamic threshold T_t(c). For a sample predicted as category c, the predicted category label and bounding box coordinates are only used as the pseudo-label if its prediction confidence is higher than the category's dynamic threshold T_t(c). Otherwise, the sample is temporarily shelved in this round of training. This step ensures that the quality of the pseudo-labels matches the current model's actual ability for each specific category, achieving fine-grained guidance of the learning process from the data selection stage.

[0130] Through the closed-loop process described above, the CPL strategy constructs an adaptive and continuously optimized data distillation mechanism in incremental learning. It not only utilizes unlabeled data macroscopically but also balances the learning progress of different categories microscopically, effectively mitigating learning biases caused by imbalanced samples or varying feature complexity. This significantly improves the overall robustness and generalization performance of the model after incremental learning.

[0131] In one embodiment, the total loss function is a weighted sum of supervised loss and unsupervised loss, expressed as:

[0132] Total Loss = A * Loss_box + B * Loss_obj + C * Loss_cls +λ* Loss_unsup;

[0133] Wherein, Loss_box is the localization loss, which adopts CIoU Loss; Loss_obj is the confidence loss, and Loss_cls is the classification loss, both of which adopt binary cross-entropy loss; Loss_unsup is the unsupervised loss calculated based on pseudo-labels; A, B, and C are the weight coefficients of localization loss, confidence loss, and classification loss, respectively, and λ is the balancing weight of unsupervised loss.

[0134] Specifically, in this embodiment, to achieve accurate and stable optimization of model parameters during incremental training, a composite total loss function is designed. This function is not a single supervised loss, but creatively combines supervised and unsupervised losses to form a unified optimization objective, thereby enabling the simultaneous utilization of both the newly added small amount of labeled data and the large amount of generated pseudo-label data. Its mathematical expression is defined as:

[0135] Total Loss = A * Loss_box + B * Loss_obj + C * Loss_cls + λ* Loss_unsup

[0136] The construction of this formula embodies clear engineering intent and algorithmic logic. The functions and design principles of each component are as follows:

[0137] The supervised loss function inherits from and optimizes the standard YOLOv5 loss function, driving model learning on samples with known correct answers (human-labeled). It consists of a weighted sum of three terms:

[0138] Location loss (Loss_box): CIoU Loss (Complete Intersection over Union Loss) is used. Unlike traditional IoU Loss, which only considers the overlap area, CIoU Loss considers the overlap area between the predicted and actual bounding boxes, the distance between their center points, and their aspect ratio. For tunnel geological prediction, the precise location of the boundaries of hazardous bodies (such as cavities and cracks) is crucial for risk assessment. CIoU Loss, through multi-dimensional optimization, can significantly improve the regression accuracy of the bounding boxes, achieving more accurate spatial positioning.

[0139] Confidence loss (Loss_obj) and classification loss (Loss_cls): Both employ binary cross-entropy (BCE). The confidence loss determines whether a target exists within the bounding box (objectivity), while the classification loss distinguishes the specific category of the target (e.g., cavity, water-rich). Using BCE instead of Softmax cross-entropy allows the model to perform multi-label classification, meaning a target may simultaneously possess multiple related attributes (e.g., a region may be simultaneously classified as "water-rich" and "fragmented"), which better reflects the ambiguity of complex geological phenomena. Weight coefficients A, B, and C (e.g., default A=0.05, B=1.0, C=0.5) are used to balance the magnitude and importance of these three losses, ensuring that the network optimizes localization, target existence judgment, and category discrimination capabilities in a balanced manner during training.

[0140] Unsupervised loss (Loss_unsup) is a key incremental term for achieving incremental learning. It is calculated based on pseudo-labels generated by the course pseudo-labeling strategy. Its calculation form is similar to Loss_box + Loss_obj + Loss_cls in supervised loss, but here the "true values" are replaced by "pseudo-labels" generated by the model itself. This loss term measures the consistency between the model's current prediction and its own past high-confidence predictions. By minimizing Loss_unsup, the model is encouraged to maintain the stability and consistency of its predictions on a large amount of unlabeled data, thereby incorporating the potential data distribution patterns and feature patterns contained in the pseudo-labels into the model parameters.

[0141] The balancing weight λ is a crucial hyperparameter. λ controls the contribution of unsupervised loss to the total loss. If λ is set too large, low-quality or inaccurate pseudo-labels may dominate the optimization direction, leading to decreased model performance or model collapse; if λ is set too small, unlabeled data cannot be effectively utilized, and incremental learning is weak. In practice, λ is usually started with a small value (e.g., 0.5) and may be dynamically adjusted based on the model's performance on the validation set or the average confidence of pseudo-labels to achieve the optimal balance between utilizing new knowledge and maintaining model stability.

[0142] In one embodiment, the formula for calculating CIoU Loss is:

[0143] CIoU Loss =

[0144] Where IoU is the intersection-union ratio between the predicted bounding box and the ground truth bounding box. Let be the Euclidean distance between the centers of the predicted bounding box and the ground truth bounding box, and let c be the length of the diagonal of the smallest closed region enclosing the predicted bounding box and the ground truth bounding box. These are the weighting coefficients. This is a parameter used to measure the consistency of the aspect ratio between the predicted bounding box and the ground truth bounding box.

[0145] Specifically, in this embodiment, the loss function is a specialized optimizer for bounding box regression tasks, and its design is far superior to the traditional IoU Loss that only considers the overlapping area. In tunnel geological advance prediction scenarios, accurate spatial positioning of geological hazard targets such as cavities and cracks is directly related to the reliability of risk assessment; therefore, a refined positioning loss is crucial. CIoU Loss achieves comprehensive optimization of the predicted bounding box by introducing three complementary geometric metrics:

[0146] Overlapping area term ( This is the basic part of the loss function. IoU (Intersection over Union) calculates the ratio of the area of ​​intersection to the area of ​​union between the predicted bounding box and the ground truth bounding box. Directly penalizing the non-overlapping conditions between the two is the core driving force behind the movement of the predicted bounding box toward the ground truth target. It ensures that the model learns the ability to achieve a basic overlap between the predicted bounding box and the target region.

[0147] Center point distance ( This optimization aligns the center points of the predicted bounding box and the ground truth bounding box. Here, ρ is the Euclidean distance between the center points of the two boxes; c is the diagonal length of the smallest bounding rectangle that can simultaneously enclose both the predicted and ground truth bounding boxes. This term is added to the loss as a penalty, and its ratio form ensures the scale invariance of the metric. In GPR images, the precise location of the target's center point often corresponds to the core region of a geological anomaly. This loss effectively ensures that the predicted bounding box not only covers the target but also accurately aligns with the target's core features, avoiding situations where the bounding boxes have large areas of overlap but the center is offset.

[0148] Aspect Ratio Consistency Item ( This section, CIoU Loss, aims to optimize the shape similarity between the predicted bounding box and the ground truth bounding box. Among them, It is a parameter specifically used to measure the consistency of the aspect ratio of two frames; This is a weighting coefficient used to balance its contribution when the overlap area is sufficiently good. This loss term allows the model to focus not only on the position and size of the bounding box during optimization, but also on its aspect ratio. For geological targets in tunnel GPR images, their morphology may exhibit specific orientations or proportions (such as layered structures or ellipsoidal cavities). This loss guides the model to learn and fit these inherent morphological priors, thereby generating bounding boxes that better match the true physical morphology of the geological targets, significantly improving the accuracy of localization and the reasonableness of the interpretation results. CIoU Loss provides a powerful and robust localization optimization criterion for deep learning-based GPR image target detection by co-optimizing three key geometric factors: overlap area, center distance, and aspect ratio.

[0149] In one embodiment, the method further includes:

[0150] The performance of the incremental optimization model is evaluated based on the confusion matrix, F1-confidence curve, and detection errors of the old and new tasks. The balance weights λ of the fixed threshold τ or unsupervised loss are then adjusted according to the evaluation results.

[0151] Specifically, in this embodiment, the method not only focuses on the incremental training process of the model, but also emphasizes the systematic, multi-dimensional performance evaluation and closed-loop optimization of the training results. The performance of the model after deployment is not static; it is affected by various factors such as the distribution of new data and the quality of pseudo-labels. Therefore, establishing a scientific evaluation system and dynamically adjusting key hyperparameters based on the evaluation results is crucial to ensuring the long-term stable operation and continuous performance improvement of the entire system. This embodiment specifically includes the following evaluation and adjustment steps:

[0152] After each round of incremental learning, the resulting incrementally optimized model needs to be comprehensively evaluated on an independent validation set, which should contain data samples from different stages (new and old tasks). The evaluation is mainly based on the following three types of metrics:

[0153] Confusion matrix analysis: By calculating and visualizing the model's confusion matrix, the classification accuracy of the model across various geological hazard categories can be intuitively quantified. Pay close attention to the off-diagonal elements of the matrix: if a large number of samples from a certain category are misclassified as other categories, it indicates that the model has not adequately learned the features of that category, possibly due to a small number of samples or complex features in that category. The confusion matrix directly reveals the structural weaknesses in the model's recognition capabilities.

[0154] F1-Confidence Curve Analysis: The F1 score, the harmonic mean of precision and recall, is a core metric for evaluating the overall detection performance of a model. By plotting the F1 score as a function of the prediction confidence threshold, the calibration quality of the model's predictions can be assessed. An ideal curve should maintain a high F1 value over a wide confidence range. If the curve shows that an acceptable F1 value is only obtained when the confidence threshold is very high, it indicates that the model's predictions generally lack confidence or contain a large number of low-confidence false positives; conversely, it indicates that the prediction confidence reflects the true accuracy well. This curve is a direct basis for adjusting the false label generation threshold τ.

[0155] Comparative analysis of detection errors on new and old tasks: This is a direct means of assessing the degree of catastrophic forgetting. Calculate the model's performance metrics, such as mean AP, on the initial dataset (or previous data) representing "old knowledge" and the newly added dataset representing "new knowledge." Compare the magnitude of the decline on both. If the model's performance improves significantly on the new data but drops sharply on the old data, it indicates severe forgetting and insufficient stability of incremental learning. The results of this analysis are an important reference for adjusting the unsupervised loss weights λ.

[0156] Based on the above quantitative evaluation results, targeted adjustments were made to the two core hyperparameters that affect the incremental learning process:

[0157] Adjusting the fixed threshold τ is primarily based on the F1-confidence curve and the confusion matrix performance of each category. If the evaluation reveals that the pseudo-label recall rate for a certain category is too low (many real samples are not generated with pseudo-labels), and this category performs poorly in the confusion matrix, it may be because the current τ value is too strict for that category. In this case, the global fixed threshold τ can be appropriately lowered (e.g., from 0.7 to 0.65) so that more pseudo-label samples can be generated for this difficult category in subsequent incremental learning, thus promoting its learning. Conversely, if it is found that a large number of introduced pseudo-labels cause a decrease in model accuracy on new data (too much noise), then it may be necessary to increase the τ value to improve the pseudo-label selection criteria.

[0158] Adjusting the balancing weight λ is primarily based on the analysis of detection errors on both new and old tasks. If the evaluation shows that the model performs well on the new task but the error increases significantly on the old task, it indicates that the current value of λ may be too high, and the unsupervised loss is too aggressively guiding the model to adapt to the new distribution, leading to the forgetting of old knowledge. In this case, λ should be appropriately reduced to decrease the weight of the unsupervised loss, allowing the supervised loss (especially historical data or its inherent features) to play a greater role in optimization and to consolidate old knowledge. Conversely, if the model adapts too slowly to new data, the performance improvement is not significant, and the old knowledge is well preserved, then λ can be increased to encourage the model to more boldly utilize unlabeled data to learn new features. Figure 3 This is the confusion matrix evaluation curve of the ground-penetrating radar tunnel geological advance prediction method based on incremental learning, where... Figure 3 (a) represents non-incremental learning. Figure 3 (b) Incremental learning, as shown in the figure, indicates that the more concentrated the diagonal distribution, the better the model performance. Figure 4 This is a numerical simulation example verification diagram of the ground-penetrating radar tunnel geological advance prediction method based on incremental learning; Figure 4 (a) and Figure 4 (c) indicates a prediction made by a non-incremental learning model; Figure 4 (b) and Figure 4 (d) represents the prediction of the incremental learning model; the higher the value, the better the model performance.

[0159] A ground-penetrating radar tunnel geological advance prediction system based on incremental learning, the system comprising:

[0160] The data acquisition and preprocessing module is used to acquire and preprocess tunnel B-scan image data collected by ground penetrating radar, and to build the initial training dataset and dynamically added datasets.

[0161] The initial model training module is used to perform supervised training on the YOLOv5-based object detection base model based on an initial training dataset with labeled parameters, and generate the initial model.

[0162] The course pseudo-label generation module is used in the incremental learning phase to dynamically calculate the dynamic thresholds of each category based on the current model's prediction results for newly added unlabeled data, and then filter and generate pseudo-labels.

[0163] The incremental learning optimization module is used to initialize the model with the parameters of the initial model, combine the newly added labeled data and the generated pseudo-label data, and perform incremental training on the model by optimizing the total loss function to update the model parameters.

[0164] The intelligent geological hazard interpretation module is used to analyze and interpret the input ground-penetrating radar B-scan images in real time using a trained incremental optimization model, and automatically identify and locate geological hazard targets within them.

[0165] Specifically, this embodiment provides a ground-penetrating radar tunnel geological advance prediction system based on incremental learning. This system is a concrete implementation of the aforementioned method at the software and hardware levels. The system is designed to be deployed on servers or high-performance computing terminals at tunnel construction sites, forming a complete closed loop from data access and model training to intelligent interpretation. Its core functional modules work collaboratively to achieve lifelong learning and adaptive evolution of the model. The specific functions and implementation methods of each module are as follows:

[0166] 1. Data Acquisition and Preprocessing Module

[0167] At the hardware level, it connects to the on-site ground-penetrating radar host via standard communication protocols (such as Ethernet and USB) to receive raw radar sampling data (common formats such as .dzt and .rd3) in real time. At the software level, this module has a built-in professional data processing pipeline: first, it calls a signal processing library (such as SciPy) to perform background removal, bandpass filtering, gain restoration, and time-depth conversion on the raw data, synthesizing the one-dimensional time series into a two-dimensional B-scan grayscale image. Subsequently, the module provides basic data augmentation functions (such as random flipping and cropping) to expand the training samples. It maintains two core databases: an labeled database stores image file paths with bounding boxes and category labels, along with their annotation information, after expert interpretation; and an unlabeled database stores all newly acquired, unprocessed raw image data in order of timestamp and tunnel mileage. This module provides standardized, structured data input for all subsequent learning processes.

[0168] 2. Initial Model Training Module

[0169] This module integrates the open-source YOLOv5 framework (based on PyTorch) and pre-sets parameters specifically for GPR images. The module loads the initial training set from the annotation database and performs the model's first supervised training. Its internal process includes configuring the network structure (e.g., selecting a YOLOv5 model), setting the optimizer (e.g., SGD), learning rate strategy (e.g., Cosine decay), and data loader. During training, the module monitors the loss function's descent curve and metrics such as mean AP on the validation set in real time until the model converges. After training, it packages the generated final weight file (e.g., model_initial.pt) and training configuration file into a model repository, providing a high-starting-point benchmark model for all subsequent incremental learning.

[0170] 3. Course pseudo-tag generation module

[0171] This module is activated when the system enters an incremental learning cycle. It first reads newly added unlabeled B-scan images in batches from the unlabeled database, then loads the currently best-performing model (the initial model in the initial stage) as the "teacher model" for forward inference, obtaining the prediction confidence for each image across all categories. Subsequently, its built-in Course Pseudo-Label (CPL) algorithm begins working: based on a preset fixed threshold τ, it calculates the "learning effect" σ_t(c) for each category, and obtains the dynamic threshold T_t(c) for each category through normalization and calculation. Finally, it filters high-confidence prediction results based on T_t(c), automatically generating pseudo-label files (with the same format as real labels but marked with "pseudo") for eligible unlabeled samples. These pseudo-labels are stored in a temporary pseudo-label cache, forming the training material for the next stage along with a small amount of newly added real labeled data.

[0172] 4. Incremental Learning Optimization Module

[0173] When this module initiates an incremental training task, it loads the model weights from the previous version as initialization parameters from the model repository, thus inheriting all historical knowledge. Next, it extracts newly added real-label data from the annotation database and reads the pseudo-label data generated in this round from the pseudo-label cache, mixing them according to a preset ratio to construct the mixed training dataset for this round. During the training loop, this module uses the aforementioned composite total loss function (Total Loss = Supervised Loss + λ * Unsupervised Loss). The key is that during backpropagation, the standard gradient is calculated for the data from the real labels, while the unsupervised loss gradient is calculated for the data from the pseudo-labels (the pseudo-labels themselves are treated as constants and do not participate in the derivative). By optimizing this total loss, the model is fine-tuned based on the original parameters. After training, the newly generated incrementally optimized model is versioned and saved back to the model repository, completing one knowledge iteration. Figure 5 This is a comparison of model predictions from a ground-penetrating radar tunnel geological advance prediction method based on incremental learning. Figure 5 (a) represents the prediction of a non-incremental learning model; Figure 5 (b) represents the prediction of the incremental learning model. The higher the value, the better the model performance.

[0174] 5. Intelligent Geological Hazard Interpretation Module

[0175] This module, as a resident inference service, continuously loads the latest incrementally optimized models from the model repository. Newly acquired GPR data from the construction site, after rapid processing by the preprocessing module, can be directly pushed to this module. The module performs real-time inference on the input B-scan images, outputting structured results, including: the category name, confidence score, and pixel-level bounding box coordinates of each detected geological hazard target in the image. This module typically also integrates visualization capabilities, displaying the detection results as highlighted boxes overlaid on the original image and generating a concise forecast report containing the hazard type, approximate size, and estimated location. This report is pushed to technical personnel via a web interface or mobile device, providing immediate and intuitive intelligent support for tunnel excavation safety decisions.

[0176] In one embodiment, the course pseudo-tag generation module specifically includes:

[0177] The learning effect estimation unit is used to count the number of effective prediction samples σ_t (c) for each category based on the current model prediction results and a fixed threshold τ.

[0178] The normalization processing unit is used to normalize the learning effect estimates σ_t(c) for each category to obtain β_t(c);

[0179] The dynamic threshold calculation unit is used to calculate the dynamic threshold for each category according to the formula T_t (c) = β_t (c) * τ;

[0180] The pseudo-label filtering unit is used to identify prediction results with a prediction confidence higher than the corresponding category dynamic threshold T_t (c) as pseudo-labels.

[0181] Specifically, in this embodiment, the course pseudo-label generation module is the core algorithm engine for realizing adaptive semi-supervised learning. Its internal logic is designed to consist of four sequentially cooperating units, forming an automated processing pipeline from predictive analysis to high-quality pseudo-label output.

[0182] The learning performance estimation unit is the starting point of this process. This unit receives batch prediction results from the current inference model, including the confidence distribution of each unlabeled B-scan image across various preset geological hazard categories. Its core task is to quantitatively assess the model's current mastery of each category. The unit internally presets a global fixed confidence threshold τ (e.g., τ=0.7) as a high-confidence benchmark. This unit iterates through all predictions, performing logical judgment and counting operations for each category c: counting the number of samples predicted by the model as belonging to that category with a highest prediction confidence exceeding the threshold τ. This count is recorded as σ_t(c), which intuitively represents the size of samples belonging to category c that the model can identify with high confidence. A higher σ_t(c) value indicates that the model has learned the features of that category more thoroughly and is more robust in its identification; conversely, a lower σ_t(c) value suggests that the target category may be a current learning challenge or that there is insufficient sample size.

[0183] The normalization unit receives the estimated learning performance values ​​σ_t(c) for each category from the previous unit. Since the absolute values ​​of σ_t(c) differ significantly across categories, they are not suitable for direct cross-category comparisons and threshold adjustments; therefore, standardization is necessary. This unit performs a simple normalization operation: finding the largest σ_t(c) value among all categories, denoted as max(σ_t). Subsequently, the normalization coefficient β_t(c) for each category is calculated. After this processing, the category with the best learning performance has a β_t(c) of 1, while the β_t(c) for the remaining categories is scaled to the (0, 1) interval. This step transforms absolute values ​​into relative ability coefficients, providing a fair benchmark for subsequent differential threshold setting.

[0184] Based on the core concepts of the course, the dynamic threshold calculation unit uses the normalized coefficient β_t(c) to calculate an adaptive pseudo-label generation threshold for different categories. This unit receives the normalized coefficient β_t(c) and a system-preset fixed threshold τ. For each category c, the calculation is performed: T_t(c) = β_t(c) * τ. This formula means that for categories with high model mastery (β_t(c) close to 1), the dynamic threshold T_t(c) approaches the strict original threshold τ, ensuring that only predictions with extremely high confidence are adopted as pseudo-labels, guaranteeing reliability. For categories with low model mastery (smaller β_t(c)), the dynamic threshold T_t(c) is correspondingly reduced, allowing more "difficult samples" belonging to that category with moderate confidence but potentially correct to participate in training, specifically strengthening the model's learning in this weak area.

[0185] The pseudo-label filtering unit is the final decision-making and output stage of this module. It receives the complete model prediction results and the category-specific thresholds T_t(c) generated by the dynamic threshold calculation unit. For each unlabeled sample, this unit checks its prediction result: determines its predicted category c, and obtains the corresponding prediction confidence p. It executes the following judgment logic: if p > T_t(c), the predicted category and bounding box information of the sample are adopted and formatted into a formal pseudo-label object (usually containing category ID, bounding box coordinates, and confidence source marker); if p ≤ T_t(c), the sample is temporarily shelved in the current training cycle. All generated pseudo-labels are output to the downstream data buffer, waiting to be used for incremental training. Through the precise collaboration of this series of units, the system achieves intelligent and differentiated mining of the value of unlabeled data.

[0186] In one embodiment, the total loss function in the incremental learning optimization module is a weighted sum of CIoU localization loss based on labeled data, binary cross-entropy confidence loss and classification loss, and unsupervised loss based on pseudo-label data.

[0187] Specifically, in this embodiment, the core of the incremental learning optimization module's model parameter update lies in its design and optimization of a highly composite total loss function. This function is not a traditional single supervised loss, but a carefully constructed weighted sum designed to simultaneously drive the model to learn from both labeled and pseudo-labeled data. Specifically, this total loss function consists of two main components:

[0188] The first part is the supervised loss based on labeled data. It directly uses and integrates proven and mature loss components from the YOLOv5 object detection framework, ensuring that the model's basic detection capabilities are accurately optimized. This includes:

[0189] CIoU localization loss is used to accurately regress the bounding box position of geological hazard targets. By comprehensively optimizing the overlapping area, center point distance and aspect ratio, it is particularly suitable for the localization needs of targets with varied shapes in GPR images.

[0190] Both the binary cross-entropy confidence loss and classification loss employ the binary cross-entropy function. The confidence loss determines whether a target entity exists within the anchor frame, while the classification loss distinguishes the concrete category of the target (e.g., cavity, water-bearing). Using binary cross-entropy enables the model to support multi-label classification, effectively representing the multi-attribute characteristics of certain complex geological phenomena.

[0191] The second part is the unsupervised loss based on pseudo-labeled data. This is a key innovative introduction for incremental learning. The calculation form of this loss term is similar to that of the supervised loss mentioned above, but its optimization objective is not the manually labeled "true value," but rather the "pseudo-labels" generated by the course pseudo-label module with dynamic confidence filtering. Minimizing this loss means encouraging the model to maintain stability and consistency in its high-confidence predictions of unlabeled data, thereby absorbing and solidifying the data distribution patterns and feature patterns contained in a large amount of unlabeled data into the network parameters.

[0192] Ultimately, the incremental learning optimization module forms a unified optimization objective by weighting and summing the two loss components according to preset weighting coefficients. The hyperparameter λ acts as a crucial "regulatory valve," balancing the strength of authoritative supervision signals from labeled data and bootstrapping learning signals from pseudo-labeled data. During training, the module minimizes this total loss function through backpropagation, thus effectively utilizing valuable newly labeled data while fully exploiting the value of massive amounts of unlabeled data, based on historical model parameters as initialization. This ultimately achieves targeted updates to model parameters and robust performance improvements.

[0193] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system 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 system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0194] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0195] It should be particularly noted that, through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, or of course, by hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0196] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A ground-penetrating radar (GPR) method for advanced geological prediction of tunnels based on incremental learning, characterized in that, Includes the following steps: S1: Obtain B-scan image data of tunnel geology collected by ground penetrating radar, and construct an initial training dataset with geological hazard labels; S2: Based on the initial training dataset, supervised training is performed on the object detection basic model built on the YOLOv5 architecture to obtain the trained initial model; S3: After the initial model is deployed, acquire the newly added B-scan image data, wherein the newly added B-scan image data includes unlabeled data and partially labeled data; S4: Employ a course pseudo-labeling strategy to process the unlabeled data and dynamically generate pseudo-labels; S5: Construct a hybrid training set using the newly added labeled data and the generated pseudo-labeled data; S6: Starting with the weights of the initial model, the model is incrementally trained based on the mixed training set. The model parameters are updated by optimizing the total loss function that combines supervised loss and unsupervised loss to obtain the incrementally optimized model. The supervised loss is calculated based on labeled data, and the unsupervised loss is calculated based on pseudo-labeled data. S7: Use the incremental optimization model to detect and identify geological hazards in the new ground-penetrating radar B-scan images, and output hazard category and location information.

2. The method according to claim 1, characterized in that, The object detection model built on the YOLOv5 architecture includes the following network structure: The input layer is used to perform Mosaic data augmentation, adaptive anchor box calculation, and adaptive image scaling on the input B-scan image. The backbone network, using the CSP-Darknet53 structure, is used to extract multi-scale features from the input image; The neck network employs a CSP-PAN structure that includes an SPPF module to fuse and enhance feature information from different scales of the backbone network. The head network is used for target bounding box regression, confidence prediction, and category classification based on the fused features.

3. The method according to claim 1, characterized in that, In steps S1 and S3, the preprocessing of the B-scan image data also includes one or more data enhancement operations such as random flipping, random cropping, and color space adjustment.

4. The method according to claim 1, characterized in that, The step of using a course pseudo-labeling strategy to process the newly added unlabeled data and dynamically generate pseudo-labels specifically includes: The newly added unlabeled data is input into the current model for prediction, and the prediction confidence scores for each category are obtained. Based on a preset fixed threshold τ, the number of samples with prediction confidence exceeding τ in each category c is counted as σ_t(c), which is used as an estimate of the learning effect of that category. The learning effect estimates σ_t(c) are normalized to obtain the normalization coefficient β_t(c); Based on the normalization coefficient β_t (c) and the fixed threshold τ, the dynamic threshold T_t (c) = β_t (c) *τ for each category c in the current learning stage is dynamically calculated. Prediction results with a confidence level higher than the corresponding class dynamic threshold T_t(c) are used as pseudo-labels for the sample.

5. The method according to claim 1, characterized in that, The total loss function is a weighted sum of supervised and unsupervised losses, expressed as: Total Loss = A * Loss_box + B * Loss_obj + C * Loss_cls +λ* Loss_unsup; Wherein, Loss_box is the localization loss, which adopts CIoU Loss; Loss_obj is the confidence loss, and Loss_cls is the classification loss, both of which adopt binary cross-entropy loss; Loss_unsup is the unsupervised loss calculated based on pseudo-labels; A, B, and C are the weight coefficients of localization loss, confidence loss, and classification loss, respectively, and λ is the balancing weight of unsupervised loss.

6. The method according to claim 5, characterized in that, The formula for calculating CIoU Loss is as follows: CIoU Loss = Where IoU is the intersection-union ratio between the predicted bounding box and the ground truth bounding box. Let be the Euclidean distance between the centers of the predicted bounding box and the ground truth bounding box, and let c be the length of the diagonal of the smallest closed region enclosing the predicted bounding box and the ground truth bounding box. These are the weighting coefficients. This is a parameter used to measure the consistency of the aspect ratio between the predicted bounding box and the ground truth bounding box.

7. The method according to claim 1, characterized in that, The method further includes: The performance of the incremental optimization model is evaluated based on the confusion matrix, F1-confidence curve, and detection errors of the old and new tasks. The balance weights λ of the fixed threshold τ or unsupervised loss are then adjusted according to the evaluation results.

8. A ground-penetrating radar tunnel geological advance prediction system based on incremental learning, characterized in that, The system for implementing the method of any one of claims 1 to 7 comprises: The data acquisition and preprocessing module is used to acquire and preprocess tunnel B-scan image data collected by ground penetrating radar, and to build the initial training dataset and dynamically added datasets. The initial model training module is used to perform supervised training on the YOLOv5-based object detection base model based on an initial training dataset with labeled parameters, and generate the initial model. The course pseudo-label generation module is used in the incremental learning phase to dynamically calculate the dynamic thresholds of each category based on the current model's prediction results for newly added unlabeled data, and then filter and generate pseudo-labels. The incremental learning optimization module is used to initialize the model with the parameters of the initial model, combine the newly added labeled data and the generated pseudo-label data, and perform incremental training on the model by optimizing the total loss function to update the model parameters. The intelligent geological hazard interpretation module is used to analyze and interpret the input ground-penetrating radar B-scan images in real time using a trained incremental optimization model, and automatically identify and locate geological hazard targets within them.

9. The system according to claim 8, characterized in that, The course pseudo-tag generation module specifically includes: The learning effect estimation unit is used to count the number of effective prediction samples σ_t (c) for each category based on the current model prediction results and a fixed threshold τ. The normalization processing unit is used to normalize the learning effect estimates σ_t (c) for each category to obtain β_t (c); The dynamic threshold calculation unit is used to calculate the dynamic threshold for each category according to the formula T_t (c) = β_t (c) * τ; The pseudo-label filtering unit is used to identify prediction results with a prediction confidence higher than the corresponding category dynamic threshold T_t (c) as pseudo-labels.

10. The system according to claim 8, characterized in that, The total loss function in the incremental learning optimization module is a weighted sum of CIoU localization loss based on labeled data, binary cross-entropy confidence loss and classification loss, and unsupervised loss based on pseudo-label data.

Images