CN-121980404-A - Advanced geological prediction method and system for ground penetrating radar tunnel based on incremental learning

CN121980404ACN 121980404 ACN121980404 ACN 121980404ACN-121980404-A

Abstract

The application discloses a geological advanced forecasting method and system for a ground penetrating radar tunnel based on incremental learning, and aims to solve the problems that dynamic newly-added data cannot be utilized, a large number of labels are relied on and disastrous forgetfulness exists in the prior art. According to the application, an incremental learning mechanism is combined with a target detection model, and by introducing a course pseudo-label strategy, the model can learn efficiently from a continuously generated GPR actual measurement data stream containing a large amount of unlabeled data, so that the prediction precision and generalization capability are continuously improved, forgetting of old knowledge is effectively relieved, continuous self-optimization can be realized by utilizing a large amount of unlabeled data included in dynamically newly added actual measurement data under the support of a small amount of labels, the memory of the existing knowledge is maintained while the model is adapted to a new environment and a new hazard type, and the intelligent tunnel geological hazard recognition and prediction with high precision, high efficiency and high generalization capability is finally realized.

Inventors

LI JINGHE
LIU LIAN
YANG TINGWEI
WANG HONGHUA
JIANG HONGLIANG
LU CHAOBO
Xiong Chunfa
Ran Mengkun

Assignees

桂林理工大学

Dates

Publication Date: 20260505
Application Date: 20260109

Claims (10)

1. The advanced geological prediction method for the ground penetrating radar tunnel based on incremental learning is characterized by comprising the following steps of: S1, acquiring B-scan image data of tunnel geology acquired by a ground penetrating radar, and constructing an initial training data set with geological hazard labeling labels; S2, performing supervision training on a target detection basic model constructed based on YOLOv architecture based on the initial training data set to obtain a trained initial model; s3, after the initial model is deployed, newly added B-scan image data is obtained, wherein the newly added B-scan image data comprises unlabeled data and part of labeled data; s4, adopting course pseudo-tag strategy to process the unlabeled data and dynamically generating pseudo tags; S5, constructing a mixed training set by using the newly added marked data and the generated data with the pseudo tag; s6, performing incremental training on the model based on the mixed training set by taking the weight of the initial model as a starting point, and updating model parameters by optimizing a total loss function combining supervision loss and non-supervision loss to obtain an incremental optimization model, wherein the supervision loss is calculated based on the marked data, and the non-supervision loss is calculated based on the data with the pseudo tag; and S7, detecting and identifying geological hazards of the new ground penetrating radar B-scan image by utilizing the incremental optimization model, and outputting hazard type and position information.
2. The method of claim 1, wherein the object detection model constructed based on YOLOv architecture, the network structure of the object detection model comprises: the input layer is used for performing Mosaic data enhancement, self-adaptive anchor frame calculation and self-adaptive picture scaling on the input B-scan image; the backbone network adopts a CSP-Darknet53 structure and is used for extracting multi-scale characteristics from an input image; the neck network adopts a CSP-PAN structure comprising an SPPF module and is used for fusing and enhancing the characteristic information of different scales from the backbone network; And the head network is used for carrying out target bounding box regression, confidence prediction and category classification based on the fused features.
3. The method of claim 1, wherein the preprocessing of the B-scan image data in steps S1 and S3 further comprises one or more data enhancement operations of random flipping, random clipping, and color space adjustment.
4. The method of claim 1, wherein the step of dynamically generating pseudo tags by processing the newly added unlabeled data using a course pseudo tag policy comprises: Inputting the newly added unlabeled data into the current model for prediction, and obtaining the prediction confidence of each category; based on a preset fixed threshold value tau, counting the number sigma_t (c) of samples with prediction confidence exceeding tau in each category c, and taking the number sigma_t (c) as learning effect estimation of the category; Normalizing each learning effect estimation sigma_t (c) to obtain a normalized coefficient beta_t (c); Dynamically calculating a dynamic threshold t_t (c) =β_t (c) τ of each category c in the current learning stage according to the normalization coefficient β_t (c) and the fixed threshold τ; and taking a prediction result with the prediction confidence higher than the corresponding category dynamic threshold T_t (c) as a pseudo tag of the sample.
5. The method of claim 1, wherein the Total Loss function Total Loss 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; The method comprises the steps of determining a Loss_box, wherein the Loss_box is a positioning Loss, the Loss_obj is a confidence Loss, the Loss_cls is a classification Loss, the Loss_cls and the Loss_box are binary cross entropy losses, the Loss_ unsup is an unsupervised Loss calculated based on a pseudo tag, A, B and C are weight coefficients of the positioning Loss, the confidence Loss and the classification Loss respectively, and lambda is balance weight of the unsupervised Loss.
6. The method of claim 5, wherein the CIoU Loss formula is: CIoU Loss = Wherein IoU is the intersection ratio of the predicted frame and the real frame, For the euclidean distance of the predicted and real frame center points, c is the diagonal length of the minimum closed region surrounding the predicted and real frames, As the weight coefficient of the light-emitting diode, Is a parameter for measuring the aspect ratio consistency of the predicted frame and the real frame.
7. The method according to claim 1, wherein the method further comprises: And evaluating the performance of the incremental optimization model based on the confusion matrix, the F1-confidence curve and the new and old task detection errors, and adjusting the fixed threshold tau or the unsupervised loss balance weight lambda according to an evaluation result.
8. A ground penetrating radar tunnel geology advanced prediction system based on incremental learning, characterized in that it is adapted to implement the method of any one of claims 1 to 7, said system comprising: The data acquisition and preprocessing module is used for acquiring and preprocessing tunnel B-scan image data acquired by the ground penetrating radar and constructing an initial training data set and a dynamic newly-added data set; the initial model training module is used for performing supervision training on the target detection basic model based on YOLOv architecture based on an initial training data set with labeling labels to generate an initial model; The course pseudo tag generation module is used for dynamically calculating dynamic thresholds of various categories by applying course pseudo tag strategies according to the prediction result of the current model on the newly added unlabeled data in the incremental learning stage, and screening and generating pseudo tags; The incremental learning optimization module is used for carrying out incremental training on the model by taking the parameters of the initial model as initialization and combining the newly added labeling data and the generated pseudo tag data and updating the model parameters through optimizing the total loss function; And the intelligent geological hazard interpretation module is used for utilizing the trained incremental optimization model to analyze and interpret the input ground penetrating radar B-scan image in real time and automatically identifying and positioning geological hazard targets in the ground penetrating radar B-scan image.
9. The system of claim 8, wherein the course pseudo tag generation module specifically comprises: The learning effect estimation unit is used for counting the effective prediction sample number sigma_t (c) of each category according to the current model prediction result and the fixed threshold tau; the normalization processing unit is used for normalizing the learning effect estimation sigma_t (c) of each category to obtain beta_t (c); A dynamic threshold calculation unit for calculating a dynamic threshold for each category according to the formula t_t (c) =β_t (c) ×τ; And the pseudo tag screening unit is used for determining a prediction result with the prediction confidence higher than the corresponding category dynamic threshold T_t (c) as a pseudo tag.
10. The system of claim 8, wherein the total loss function in the incremental learning optimization module is a weighted sum of CIoU localization losses, binary cross entropy confidence losses, and classification losses based on labeling data, and unsupervised losses based on pseudo tag data.

Description

Advanced geological prediction method and system for ground penetrating radar tunnel based on incremental learning The application relates to the technical field of safe and environment-friendly construction detection, in particular to a geological advanced forecasting method and system for a ground penetrating radar tunnel based on incremental learning. Background By the time of recent years, the number of established tunnels and total mileage in China are all in the world, and a high-level construction situation is still kept in the future. However, the risks of geological disasters such as collapse, water flooding, lining back holes and the like in tunnel construction and operation are increasingly remarkable, and great personnel and economic losses are caused. Therefore, accurate and efficient advanced geological forecast is performed in the construction period of the tunnel, and the method is a core link for disaster prevention and safety guarantee. Ground penetrating radar (group PENETRATING RADAR, GPR) is used as a nondestructive geophysical detection means, and is one of the mainstream technologies for advanced prediction of tunnel geology, wherein poor geological bodies such as rock mass structures, water-bearing zones, karst caves and broken zones in front of a tunnel face can be effectively detected by transmitting high-frequency electromagnetic waves and receiving reflected signals from an underground medium. The B-scan image generated by the method is a visual reflection of electromagnetic wave time-space response. However, interpretation of GPR data is highly dependent on the expertise of the interpreter, and presents inherent challenges of being subjective, inefficient, difficult to identify complex interfering signals (i.e., the "multi-resolvable" problem). Particularly in tunnel environments with complex medium conditions and changeable target body forms, the accuracy and consistency of manual interpretation are difficult to ensure. In recent years, machine learning, particularly deep learning technology, has made breakthrough progress in the fields of image recognition and target detection, and provides a new approach for intelligent interpretation of GPR data. Some researches try to apply convolutional neural network, faster R-CNN, yolo and other models to automatic identification of hazard targets in GPR images, and obtain performance superior to that of the traditional method. However, existing depth learning based GPR interpretation schemes typically suffer from the following key drawbacks: Firstly, model training is static and insufficient in data utilization, and most of the existing methods adopt a 'one-time' training mode. That is, prior to model deployment, a fixed, finite (typically containing a large amount of numerical simulation data and a small amount of measured data) data set is used for training and validation, with the model parameters then frozen. Tunnel engineering is a typical dynamic continuous process, and as tunneling advances, a large amount of new GPR measured data reflecting different surrounding rock conditions is continuously generated. The dynamic newly-added data contains rich field information, but the traditional model lacks continuous learning capability, and cannot be utilized to optimize and update the model, so that the generalization capability of the model is limited, and the model is difficult to adapt to the forecasting requirements under long-distance and changeable geological conditions. Secondly, data annotation bottleneck, supervised learning is a current mainstream mode, and the performance of the data annotation bottleneck is seriously dependent on large-scale and high-quality annotation data. However, accurate geological hazard labeling of GPR images requires geophysical specialists to consume a lot of time and effort, and is costly. In engineering practice, the newly acquired data often cannot be marked in time, so that a large amount of unmarked data is formed. The traditional method is either abandoned, so that data waste is caused, or the model is retrained after the manual labeling is completed, so that the process is tedious, and the requirement of construction on timeliness cannot be met. Finally, the problem of catastrophic forgetting is that if the model is simply retrained after mixing new and old data, new knowledge can be learned, but old knowledge learned previously is often seriously covered or forgotten, resulting in the dramatic decrease of the performance of the model on the data already seen, which is not acceptable in a dynamic incremental engineering scenario. Incremental learning (INCREMENTAL LEARNING), or continuous learning (Continual Learning), aims to enable models to learn new knowledge continuously like humans from a continuously arriving data stream while preserving memory of learned knowledge as much as possible. This provides a theoretical framework for solving the dynamic adaptation problem in GPR intelligence i