Tunnel Lining GPR Defect Detection Using Self-Supervised Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current ground-penetrating radar (GPR) image recognition methods based on convolutional neural networks (CNN) require extensive manual labeling of training samples, leading to high time and economic costs, and the generated training samples are not real measurement data, resulting in low recognition accuracy.
Innovation Solution
A self-supervised learning method using unlabeled GPR data for pre-training a feature extraction network through contrastive learning, followed by fine-tuning with labeled data to construct a convolutional neural network model for tunnel lining quality detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual labeling of GPR detection images is performed to prepare training sample sets for CNN-based deep learning, then the recognition accuracy can be improved, but the time cost and economic cost increase significantly
Solution Approach 1:
The patent applies self-supervised learning where the model learns to recognize features from unlabeled GPR images itself, without requiring manual annotation. The system automatically identifies patterns and features in the data, eliminating the need for human labelers while maintaining high recognition accuracy for tunnel lining defects
Solution Approach 2:
The patent performs pre-training on large amounts of unlabeled data before fine-tuning on small labeled datasets. This preliminary learning phase allows the model to acquire general features from abundant unlabeled GPR images, reducing the subsequent need for extensive manual labeling while improving final recognition performance
2Measurement precision
If more training samples are collected to improve recognition accuracy, then the model performance can be enhanced, but the data labeling cost and time consumption increase
Solution Approach 1:
The model performs self-supervised learning on unlabeled data, automatically discovering features and patterns without human intervention in the labeling process. This enables utilization of large datasets without proportional increases in labeling costs
Solution Approach 2:
The patent uses a small amount of labeled data for fine-tuning after extensive pre-training on unlabeled data. This partial labeling approach achieves high accuracy without requiring complete labeling of all training samples, improving labeling efficiency
3Quantity of substance
If GPR detection data is manually labeled by different personnel, then training samples can be prepared, but interpretation errors are introduced due to experience differences
Solution Approach 1:
The self-supervised learning framework eliminates human labelers from the process, allowing the model to learn directly from raw GPR images without subjective interpretation. This ensures consistent feature extraction across all training samples regardless of operator experience levels
Solution Approach 2:
The patent uses data augmentation techniques to generate synthetic training samples through transformations of existing unlabeled data. This creates additional training examples without introducing new human interpretation errors, maintaining label consistency while increasing sample quantity
Data Source
AI summary
A method for a quality detection of tunnel lining through ground-penetrating radar based on self-supervised learning. In the method, a grayscale image of the tunnel to be detected is obtained, the grayscale image is then input into a trained feature recognition model, to obtain a feature atlas is corresponding to the grayscale image, and then a quality recognition result of the tunnel to be detected is determined according to the feature atlas. In the present application, the feature recognition model is obtained by means of the self-supervised learning according to an unlabeled image set and a labeled image set, and the unlabeled image set is directly utilized to train the recognition model, which not only improves the efficiency of training the tunnel recognition model, but also improves the accuracy of tunnel quality detection results.