Research On Automatic Interpretation Algorithm Of GPR Image In Highway Tunnel Detection

With the development of highway construction, the tunnel scale constantly increases accordingly. However, the tunnels in-service often suffer from various damages, such as, voids, cracks and fissures, etc, which not only affects the normal operation of the economy, but also poses a threat to people’s safety. Therefore, tunnel inner inspection is considered to be a necessary and significant work. Ground penetrating radar (GPR), because of its fast and nondestructive characteristics, is applied to perform tunnel inner inspection, which gives a global view of tunnel structural situation and provides reference and evidence for tunnel maintenance. Interpretation of GPR data is usually done manually, but it is time-consuming and error-prone processes. Consequently, a fast and accurate automatic method for tunnel damages detection is proposed in this paper. The main contents are as following:1. Tunnel damages simulation basing FDTD:Firstly, simulations are made to explore the investigation depth and precision of different antenna, which provides frequency choice with reference. And then, in order to obtain the prior information and characteristics of different damages in GPR image, Finite-difference Time-domain method is used to model the tunnel structure and probable damage inside tunnel, which brings convenience for damage comprehension and identification. The simulation is not only conducive to train expert for data interpretation, but also provides reference for damage analysis and data interpretation of measurement and guide the tunnel inspection work effectively.2. Damage inside tunnel detection and classification:Because that strong clutter and noise in GPR image, auto focus of hyperbola and representativly feature extraction make it difficult to perform data interpretation, data pre-processing basing on simulated and measured data are performed by filters designed in time domain, frequency domain and subspace domain. On this basis. migration method and time varying gain are utilized to focus hyperbola caused by target (rebar) and damages and enhance the damages signal. respectively, and then. representative features are extracted. Therefore. these pre-processing methods effectively solve the difficulties in enhancing signal-noise ratio, auto focus of hyperbola and representatively feature extraction for data interpretation. At last, a learning machine support vector machine (SVM) basing simulated and measured data is used to perform the detection and classification automatically.The mean accuracy of correct mlti-defect classification basing simulated data is94.53%. The goodness of fit between two class classification using SVM and expert empirical interpretation results is84.6%. Further central distance analysis of features indicates that SVM predicted result is correct and the reason why there exists conflict is that expert empirical interpretation results include erroneous judgement and is not enough accurate. Consequently, SVM predicted result not only could instruct expert to perform data interpretation, but also revise it. And it can shorten the interpretation period and avoid the multiple solutions as well.