Nondestructive testing method for tunnel lining defects

Abstract

The invention discloses a tunnel lining defect non-destructive testing method, and relates to the technical field of non-destructive testing. During operation of a system, an image acquisition device is used for acquiring continuous image sequences of a tunnel lining, spatial registration is realized based on feature point matching between adjacent image sequences, an initial multi-source image set is obtained, the initial image set is preprocessed, and the initial multi-source image set is obtained; obtaining an enhanced multi-source image sequence, carrying out feature extraction and defect detection, extracting hidden crack edge features by adopting a multi-scale convolutional network, inhibiting dynamic interference in combination with an image-to-image time sequence difference detection model, generating a segmentation image of a hidden crack defect region, and carrying out segmentation on the hidden crack defect region; the method comprises the following steps: identifying hidden cracks, acquiring width and depth indication characteristics and position parameters of the hidden cracks through crack skeletonization mapping, inputting the width and depth indication characteristics and position parameters into a multi-task learning model, constructing a defect severity index based on the identified hidden crack parameters, performing health assessment on the evolution state of the hidden cracks in combination with time sequence trend prediction, and automatically generating maintenance priorities, risk levels and maintenance suggestions.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of nondestructive testing, in particular to a tunnel lining defect nondestructive testing method. BACKGROUND

[0002] As a key component to ensure the safety and service life of the tunnel structure, the tunnel lining is inevitably affected by environmental factors, vehicle load, water penetration and temperature and humidity changes during long-term operation, thereby causing various diseases such as cracks, hollowing and peeling. The traditional tunnel lining defect detection method mainly relies on manual inspection or single sensor means, which has the problems of low efficiency, strong subjectivity and difficulty in identifying hidden defects. Especially in the tunnel operation state, limited by low illumination, vehicle interference and wet stains, cracks often exist in the form of hidden cracks, which are small in width and fuzzy in boundary, and often cannot be effectively identified by conventional detection methods. Therefore, how to realize nondestructive testing of lining hidden cracks in the tunnel in operation has become a technical problem to be solved.

[0003] The existing tunnel lining defect nondestructive testing system mainly relies on laser scanning or ultrasonic detection methods, which can obtain geometric data or internal reflection characteristics, but has the defects of high hardware cost, complex construction layout, large amount of data and insufficient real-time performance. At the same time, although the detection method based on single modal image has the advantages of low cost and flexible deployment, it is easily affected by complex environmental factors such as vehicle light glare, dust interference, water leakage and insufficient illumination during tunnel operation, resulting in low image signal-to-noise ratio, fuzzy crack edge and insufficient accuracy of detection results. This deficiency makes the system have obvious defects in the early detection of hidden cracks, and it is difficult to realize rapid warning of potential diseases.

[0004] The above-mentioned deficiencies of the present situation are mainly caused by the superposition of dynamic interference factors in the tunnel operation environment and the concealment of lining defects. On the one hand, the illumination and aerosol dust generated by the running vehicle make the image acquisition prone to overexposure, speckle and noise; on the other hand, wet stains and temperature changes cause low-contrast areas on the lining surface, thereby masking the boundary features of the cracks. If hidden cracks cannot be detected and identified in time, they will easily develop into structural cracks over time, causing local peeling of the lining, increased leakage and even instability of the lining, etc. This not only shortens the service life of the tunnel, but also may cause major hidden dangers to the operation safety. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application provides a tunnel lining defect nondestructive testing method, which solves the problems mentioned in the background art.

[0006] To achieve the above purpose, the present application realizes the following technical scheme: a tunnel lining defect nondestructive testing method, comprising the following steps:

[0007] S1: In the tunnel operation state, a continuous image sequence of the tunnel lining is obtained by using an image acquisition device. The image acquisition process includes running interference factors, including vehicle light glare, low illumination, wet stains, and dust particles; spatial registration is achieved based on feature point matching between adjacent image sequences to obtain an initial multi-source image set;

[0008] S2: The initial image set is preprocessed, the vehicle light glare effect is eliminated based on a light suppression and dynamic range enhancement model, the dust particle noise is suppressed using a convolutional neural network adaptive denoising model, and the low-contrast area affected by the wet stain is compensated by fusing the thermal infrared image and the visible light image to obtain an enhanced multi-source image sequence;

[0009] S3: Feature extraction and defect detection are performed on the multi-source image sequence, a multi-scale convolutional network is used to extract hidden crack edge features, a time series difference detection model is combined to suppress dynamic interference, a segmentation map of the hidden crack defect area is generated, and the width, depth, and location parameters of the hidden crack are obtained by crack skeleton mapping;

[0010] S4: The defect segmentation result is input into a multi-task learning model, and the hidden crack defect category, spatial scale, and center offset are output by a joint classification and regression network;

[0011] S5: Based on the identified hidden crack parameters, a defect severity index is constructed, and a time series trend prediction is combined to evaluate the evolution state of the hidden crack, and a maintenance priority, risk level, and maintenance suggestion are automatically generated.

[0012] Preferably, high-definition cameras are used to acquire visible light images, while infrared thermal imaging images and short-baseline binocular images are simultaneously collected to form a multi-source image set of visible light-infrared-depth;

[0013] Based on the cross-modal registration method of image-to-image, the infrared temperature distribution is aligned with the visible light texture to compensate for the early features of the hidden crack caused by wet stains;

[0014] A sparse depth map is generated from the binocular image sequence, which is jointly input into the detection model with the visible light image to realize three-dimensional scale recovery of the hidden crack area.

[0015] Preferably, when performing adaptive denoising on the collected images, first identify the noise type as Gaussian noise, impulse noise, or dust scattering noise, then use the CNN denoising model to perform targeted noise reduction on different noise patterns, and calculate the image clarity enhancement factor Ec to dynamically adjust the contrast and brightness of the image, making the hidden crack edge prominent and ensuring the accuracy of subsequent crack skeleton extraction.

[0016] Preferably, the image clarity enhancement factor E c is calculated by the following formula:

[0017] ;

[0018] E c represents the image sharpness enhancement factor, represents the local gray standard deviation, C d represents the image detail complexity, represents the overall average gray level of the image, represents the minimum value, avoiding the denominator being 0;

[0019] Preferably, different size crack and hole features are extracted by using a multi-scale CNN for hierarchical feature extraction on the preprocessed image, and the crack skeleton point set (xi, yi) is serialized in the crack skeleton mapping process, and the physical length Lc of the hidden crack is obtained by the following formula:

[0020] ;

[0021] L c represents the physical length of the crack, and N represents the number of crack skeleton points, represents the mapping ratio of pixels to physical space;

[0022] The crack skeleton points are spatially mapped to obtain the real length of the crack in the physical space.

[0023] Preferably, a deep learning model is constructed to simultaneously complete defect classification, positioning and size estimation, the defect classification includes crack, hollow and falling, transfer learning is used, the model pre-trained on a large-scale data set of road or building cracks is transferred to the tunnel lining scene, and a small amount of tunnel lining defect data is used for fine tuning, the image input, feature extraction, defect recognition and positioning tasks are unified in an end-to-end deep learning model, and the original image is directly input into the network to automatically complete detection and recognition.

[0024] Preferably, a quantitative health assessment value is generated based on the detected defect result, and a defect severity index D s is calculated to quantify the current defect severity, and based on the defect severity index D s , a reinforcement learning algorithm is used to predict the future defect development trend, and a defect development trend prediction value D p (t+Δt) is calculated, different defects are assigned repair priorities, defects are visualized in the three-dimensional model through BIM and AR technologies, and the system automatically generates an evaluation report and repair recommendations.

[0025] Preferably, the defect severity index D s is calculated by the following formula:

[0026] ;

[0027] D s represents a defect severity index, L c represents a true length of a crack, L t represents a reference length of a tunnel single-segment lining, W c represents an average width of a crack, W t represents a maximum width allowed in design, ΔT represents a temperature anomaly value of a defect area, T ref represents a normal temperature of a lining, and α, β and γ represent weighting factors;

[0028] a defect development trend prediction value D p (t+Δt) is calculated by the following formula:

[0029] ;

[0030] D p (t+Δt) represents a predicted defect severity at a future time t+Δt, D s (t) represents a defect severity index at a current time, λ represents a defect growth rate factor, and Δt represents a prediction time interval, represents a rate of change of a defect severity index with respect to time, represents a trend correction coefficient.

[0031] Preferably, based on the defect severity and the prediction result, a maintenance priority, a risk level and a maintenance suggestion are automatically generated, and the results are combined with a three-dimensional BIM model to display the defect position and severity through an augmented reality (AR) mode.

[0032] A preset severity threshold Ts and a growth threshold Tg are provided.

[0033] If a current defect severity index D s is greater than or equal to a preset severity threshold Ts, and a predicted defect severity index D p has a growth rate greater than or equal to a preset growth threshold Tg within a future time interval Δt, it is determined that the lining area corresponding to the defect is a high-priority maintenance area; otherwise, it is determined that the lining area corresponding to the defect is a low-priority maintenance area.

[0034] Preferably, a risk threshold Tr and an allowable width threshold Wt are provided to form a future trend evaluation strategy.

[0035] If a predicted defect severity index D p is greater than or equal to a preset risk threshold Tr within a future time interval Δt, and a current crack width W cIf the difference is greater than or equal to a preset allowable width threshold Wt, the lining area corresponding to the defect is marked as a high-risk level, and a mandatory maintenance suggestion is triggered; otherwise, the lining area corresponding to the defect is marked as a general-risk level, and a routine inspection suggestion is generated.

[0036] The application provides a tunnel lining defect nondestructive detection method, which has the following beneficial effects:

[0037] (1) When the system is running, a continuous image sequence of the tunnel lining is acquired by using an image acquisition device, spatial registration is realized based on feature point matching between adjacent image sequences, an initial multi-source image set is obtained, the initial image set is preprocessed, an enhanced multi-source image sequence is obtained and feature extraction and defect detection are performed, multi-scale convolution network is used to extract hidden crack edge features, an image-to-image time series difference detection model is combined to suppress dynamic interference, a segmentation map of the hidden crack defect area is generated, the width, depth indication features and position parameters of the hidden crack are obtained through crack skeleton mapping, a multi-task learning model is input, a defect severity index is constructed based on the identified hidden crack parameters, and the evolution state of the hidden crack is health evaluated in combination with time series trend prediction, maintenance priority, risk level and maintenance suggestion are automatically generated.

[0038] (2) The application forms a complete tunnel lining defect nondestructive detection process through the organic combination of the six modules. Starting from continuous image acquisition under the tunnel operation state, the preprocessing module is used to remove the influences of light interference, noise pollution and low-contrast wet stains, the feature extraction and detection module is used to realize accurate segmentation of hidden cracks, and finally the deep learning module is used to complete defect classification and parameter regression. The whole system not only realizes early identification of hidden cracks in the tunnel lining, but also completes key tasks such as defect parameterization description, severity index calculation and trend prediction, thereby providing complete data support for subsequent operation and maintenance decisions.

[0039] (3) Compared with the existing detection methods relying on artificial inspection or single sensors, the application realizes fusion of multi-source image data and stable detection under dynamic interference, and avoids the defects of low image signal-to-noise ratio and blurred crack edges in single-mode detection. The preprocessing module overcomes the problems of running environment such as car light glare, dust interference and wet stain covering through light suppression, CNN adaptive denoising and cross-modal fusion. Meanwhile, the deep learning module uses a multi-task learning framework to unify classification, positioning and size estimation in an end-to-end network, thereby improving the automation degree and accuracy of detection. These improvements enable the system to maintain high reliability and high precision in complex operating environments, thereby breaking through the bottleneck of real-time performance and adaptability of the existing technology.

[0040] (4) By introducing the defect severity index and the trend prediction model in the evaluation and feedback module, the present application realizes the leap from "detection" to "prediction and decision", not only quantifying the current lining health status, but also dynamically predicting the defect development trend, automatically generating maintenance priority and risk level. At the same time, combining BIM and augmented reality display means, the detection results can be intuitively visualized in three-dimensional space, greatly improving the cognitive efficiency of engineers on defect distribution and severity. In summary, the present application has significantly improved in detection accuracy, environmental adaptability, real-time performance and result visualization, realizing the transformation from traditional static detection to intelligent, visual and predictive detection, and comprehensively enhancing the safety and scientificity of tunnel lining operation and maintenance. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 It is a tunnel lining defect nondestructive testing method step schematic diagram of the present application;

[0042] Figure 2 It is a tunnel lining defect nondestructive testing method flowchart of the present application. DETAILED DESCRIPTION

[0043] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0044] Embodiment 1

[0045] The present application provides a tunnel lining defect nondestructive testing method, please refer to Figure 1 , comprising the following steps:

[0046] S1: In the tunnel operation state, the continuous image sequence of the tunnel lining is obtained by using the image acquisition device, and the image acquisition process contains running interference factors, including car light glare, low illumination, water leakage wet stain and dust particles; the spatial registration is realized based on the feature point matching between adjacent image sequences, and the initial multi-source image set is obtained;

[0047] S2: The initial image set is preprocessed, the car light glare influence is eliminated based on the illumination suppression and dynamic range enhancement model, the dust particle noise is suppressed by using the convolutional neural network adaptive denoising model, and the fusion compensation of the low contrast area affected by the wet stain is carried out by using the thermal infrared image and the visible light image, and the enhanced multi-source image sequence is obtained;

[0048] S3: Perform feature extraction and defect detection on the multi-source image sequence, extract crack edge features using a multi-scale convolutional network, suppress dynamic interference using an image-to-image temporal difference detection model, generate a segmentation map of the hidden crack defect area, and obtain the width, depth indication features and position parameters of the hidden crack through crack skeleton mapping;

[0049] S4: Input the defect segmentation result into a multi-task learning model, output the hidden crack defect category, spatial scale and center offset through a joint classification and regression network;

[0050] S5: Based on the identified hidden crack parameters, construct a defect severity index, and combine time series trend prediction to evaluate the evolution state of the hidden crack, automatically generate maintenance priority, risk level and maintenance suggestions.

[0051] In this embodiment, a continuous image sequence of the tunnel lining is obtained by using an image acquisition device, spatial registration is realized based on feature point matching between adjacent image sequences, an initial multi-source image set is obtained, the initial image set is preprocessed, an enhanced multi-source image sequence is obtained and feature extraction and defect detection are performed, a multi-scale convolutional network is used to extract crack edge features, an image-to-image temporal difference detection model is used to suppress dynamic interference, a segmentation map of the hidden crack defect area is generated, and the width, depth indication features and position parameters of the hidden crack are obtained through crack skeleton mapping, input into a multi-task learning model, based on the identified hidden crack parameters, construct a defect severity index, and combine time series trend prediction to evaluate the evolution state of the hidden crack, automatically generate maintenance priority, risk level and maintenance suggestions.

[0052] Embodiment 2

[0053] This embodiment is an explanation and description in Embodiment 1, please refer to Figure 1 , specifically: obtain visible light images through a high-definition camera, simultaneously collect infrared thermal imaging images and short-baseline binocular images, form a multi-source image set of visible light-infrared-depth;

[0054] Based on the cross-modal registration method of image-to-image, align the infrared temperature distribution with the visible light texture, realize the early feature compensation of hidden cracks caused by wet stains;

[0055] Generate a sparse depth map through a binocular image sequence, input the visible light image into a detection model, realize three-dimensional scale recovery of the hidden crack area.

[0056] When adaptively denoising the collected image, first, the noise type is identified as Gaussian noise, impulse noise or dust scattering noise, then a CNN denoising model is used to perform targeted denoising on different noise modes, and an image definition enhancement factor Ec is calculated to dynamically adjust the contrast and brightness of the image, so that the hidden crack edge is highlighted, and the accuracy of subsequent crack skeleton extraction is ensured.

[0057] In this embodiment, the visible light image is obtained by a high-definition camera, and the multi-source image set of visible light-infrared-depth is formed by combining infrared thermal imaging and short-baseline binocular images, which can realize all-around capture of hidden crack defects in the tunnel operating environment; wherein, based on the cross-modal registration method from image to image, the infrared temperature distribution and the visible light texture are accurately aligned, the low-contrast area caused by wet stains is effectively compensated, and the saliency of the early features of hidden cracks is improved; at the same time, the sparse depth map generated by the binocular image sequence is combined with the visible light image to input the detection model, which can restore the three-dimensional spatial scale of the hidden crack area and enhance the quantitative description ability of the crack depth and position; further, in the image preprocessing process, through adaptive identification and targeted CNN denoising processing of the noise type, and combined with dynamic contrast and brightness adjustment of the image definition enhancement factor Ec, the hidden crack edge can be highlighted, and the influence of factors such as headlight glare, dust interference and wet stain cover during operation can be suppressed, thereby ensuring the accuracy of crack skeleton extraction. Overall, the present application effectively improves the accuracy, robustness and early warning ability of hidden crack detection of tunnel lining.

[0058] Embodiment 3

[0059] This embodiment is an explanation and description in embodiment 1, please refer to Figure 1 , specifically: the image definition enhancement factor E c is calculated by the following formula:

[0060] ;

[0061] In the formula, E c represents the image definition enhancement factor, represents the local gray scale standard deviation, C d represents the image detail complexity, represents the overall average gray scale of the image, represents the minimum value, to avoid the denominator being 0;

[0062] By using a multi-scale CNN to extract features of different sizes of cracks and cavities from the preprocessed image, the hidden crack skeleton point set (xi, yi) is sequentially processed in the crack skeleton mapping process, and the physical length Lc of the hidden crack is obtained by the following formula:

[0063] ;

[0064] In the formula, L c represents the physical length of the crack, N represents the number of crack skeleton points, represents the mapping ratio of pixels to physical space;

[0065] The crack skeleton points are spatially mapped to obtain the real length of the crack in the physical space.

[0066] By constructing a deep learning model, defect classification, positioning and size estimation are completed at the same time. The defect classification includes cracks, hollows and peeling. Using transfer learning, the model pre-trained on a large-scale dataset of road or building cracks is transferred to the tunnel lining scene, and a small amount of tunnel lining defect data is used for fine tuning. The image input, feature extraction, defect recognition and positioning tasks are unified in an end-to-end deep learning model and executed. The original image is directly input into the network, and the detection and recognition are automatically completed.

[0067] In this embodiment, by introducing the calculation of the image sharpness enhancement factor Ec, the comprehensive measurement of the local gray scale, the detail complexity and the overall mean of the image is realized. The dynamic balance of contrast and brightness can be realized in the preprocessing stage, thereby effectively improving the resolution between the hidden crack edge and the background and ensuring the reliability of subsequent feature extraction. Further, multi-scale convolutional neural networks are used to extract hierarchical features of cracks and cavities of different sizes, and the real physical length Lc of the hidden crack is calculated by combining the crack skeleton mapping formula, so that the detection result not only stays in the recognition at the pixel level, but also can realize the quantitative representation at the physical scale, and the engineering applicability of defect detection is enhanced. At the same time, the deep learning model realizes the unified processing of classification, positioning and size estimation through a multi-task learning structure, and combines a transfer learning strategy to transfer the knowledge of a large-scale dataset to the tunnel lining scene, effectively alleviating the problem of insufficient tunnel defect samples. Overall, the present application not only improves the precision and stability of tunnel hidden crack detection, but also realizes quantitative, automatic and intelligent recognition of defects, which is significantly superior to existing technical means relying on manual interpretation or single image features.

[0068] Embodiment 4

[0069] This embodiment is an explanation and description in embodiment 1, please refer to Figure 1 , in particular: based on the detected defect result, a quantitative health assessment value is generated, and a defect severity index D s is calculated to quantify the current defect severity, based on the defect severity index D s , a reinforcement learning algorithm is used to predict the future defect development trend, and a defect development trend prediction value D pThe defect severity index D(t+Δt) is assigned a repair priority, the defects are visualized in the three-dimensional model through BIM and AR technology, and the system automatically generates an evaluation report and repair recommendations.

[0070] Defect severity index D s The defect severity index D(t+Δt) is calculated by the following formula:

[0071] ;

[0072] In the formula, D s represents the defect severity index, L c represents the true length of the crack, L t represents the reference length of the single segment lining of the tunnel, W c represents the average width of the crack, W t represents the maximum width allowed by the design, ΔT represents the temperature anomaly value of the defect area, T ref represents the normal temperature of the lining, and α, β and γ represent weighting factors.

[0073] Defect development trend prediction value D p The defect development trend prediction value D(t+Δt) is calculated by the following formula:

[0074] ;

[0075] In the formula, D p (t+Δt) represents the defect severity predicted at future time t+Δt, D s (t) represents the defect severity index at the current time, λ represents the defect growth rate factor, and Δt represents the prediction time interval, represents the rate of change of the defect severity index with time, represents the trend correction coefficient.

[0076] In this embodiment, the defect severity index D s The present application can convert multi-dimensional parameters such as crack length, width and temperature anomaly into a unified quantitative index, thereby realizing accurate characterization of the current lining health status. Further, the defect development trend prediction value D pThe defect severity evolution over time is modeled and predicted by combining the reinforcement learning algorithm, which not only can realize the early prediction of potential risks, but also can assign reasonable maintenance priorities to different types of defects, and significantly improve the scientificity and foresight of maintenance decision-making. Meanwhile, the detection results are combined with the BIM three-dimensional model and augmented reality AR technology, which can intuitively display the defect position, severity and future trend in a visual way, so that the maintenance personnel can quickly understand and develop corresponding measures. In summary, the present application greatly improves the overall effect of tunnel lining crack detection in accuracy, operability and intelligent level while realizing quantitative evaluation, trend prediction and visual display.

[0077] Embodiment 5

[0078] This embodiment is an explanation and illustration in embodiment 1, please refer to Figure 1 , specifically: based on the defect severity and prediction results, automatically generate maintenance priority, risk level and maintenance suggestion, and combine the results with the three-dimensional BIM model to display the defect position and severity through augmented reality AR;

[0079] preset severity threshold Ts and growth threshold Tg;

[0080] if the current defect severity index D s is greater than or equal to the preset severity threshold Ts, and the growth rate of the predicted defect severity index D p within the future time interval Δt is greater than or equal to the preset growth threshold Tg, then the lining area corresponding to the defect is determined as a high-priority maintenance area; otherwise, the lining area corresponding to the defect is determined as a low-priority maintenance area.

[0081] preset risk threshold Tr and allowable width threshold Wt to form a future trend evaluation strategy:

[0082] if the predicted defect severity index D p within the future time interval Δt is greater than or equal to the preset risk threshold Tr, and the current crack width W c is greater than or equal to the preset allowable width threshold Wt, then the lining area corresponding to the defect is marked as a high-risk level, and a mandatory maintenance suggestion is triggered; otherwise, the lining area corresponding to the defect is marked as a general risk level, and a routine inspection suggestion is generated. Based on the determination results of high priority and high risk level, the system automatically integrates the spatial distribution characteristics and historical maintenance data to generate a hierarchical disposal work order; for the lining area that meets the high-priority and high-risk conditions at the same time, it is pushed to the maintenance scheduling platform and the emergency response process is started to ensure the priority of resource allocation; all suggestions are accompanied by confidence evaluation and uncertainty analysis to support decision optimization.

[0083] In this embodiment, the tunnel lining defect is graded and evaluated by the defect severity index D s and the double determination mechanism of the predicted result D p The application can realize grading evaluation and intelligent decision of tunnel lining defects. The preset severity threshold Ts and growth threshold Tg are used to distinguish the maintenance priority of defects, so that the system can automatically determine the high-priority and low-priority areas according to the current state and future evolution trend of the defects; at the same time, the preset risk threshold Tr and the allowable width threshold Wt are used to form a trend risk evaluation strategy to realize the forced maintenance determination of high-risk cracks. Further, when the high-priority and high-risk levels are both established, the application can automatically integrate the defect spatial distribution characteristics and historical maintenance data to generate a grading disposal work order and push the result to a maintenance scheduling platform to start an emergency response process, thereby ensuring that limited resources are preferentially allocated to the most dangerous lining area. In addition, the system outputs decision suggestions with confidence evaluation and uncertainty analysis to ensure that maintenance personnel can fully understand the reliability of the detection results and effectively avoid misjudgment and missed judgment. Compared with the existing methods relying on artificial experience or single threshold determination, the application realizes an intelligent closed loop from detection-evaluation-decision-scheduling, significantly improving the scientificity, safety and emergency response efficiency of tunnel maintenance.

[0084] Although embodiments of the application have been shown and described, it is to be understood that the application is not limited to the details of the foregoing embodiment, and that various changes in form and details can be made without departing from the spirit and scope of the application, which is defined by the following claims and their equivalents.

Claims

1. A method of non-destructive testing of defects in a tunnel lining, characterized in that: The method comprises the following steps: S1: in the tunnel operation state, the image acquisition device is used to obtain a continuous image sequence of the tunnel lining, and the image acquisition process contains running interference factors, including vehicle light glare, low illumination, wet stains and dust particles; spatial registration is realized based on feature point matching between adjacent image sequences, and an initial multi-source image set is obtained; S2: the initial image set is preprocessed, the vehicle light glare influence is eliminated based on a light suppression and dynamic range enhancement model, a convolutional neural network adaptive denoising model is used to suppress dust particle noise, and a visible light image and a thermal infrared image are fused and compensated for a low-contrast area affected by wet stains, to obtain an enhanced multi-source image sequence; S3: feature extraction and defect detection are performed on the multi-source image sequence, a multi-scale convolutional network is used to extract hidden crack edge features, a time series difference detection model is combined to suppress dynamic interference, a segmentation map of the hidden crack defect area is generated, and the width, depth indication features and position parameters of the hidden crack are obtained through crack skeleton mapping; S4: the defect segmentation result is input into a multi-task learning model, and the hidden crack defect category, spatial scale and center offset are output through a joint classification and regression network; S5: a defect severity index is constructed based on the identified hidden crack parameters, and the evolution state of the hidden crack is health evaluated in combination with time series trend prediction, to automatically generate a maintenance priority, a risk level and a maintenance suggestion.

2. A method of non-destructive testing of defects in a tunnel lining according to claim 1, characterized in that: High-definition cameras are used to obtain visible light images, infrared thermal imaging images and short-baseline binocular images are collected at the same time, and a multi-source image set of visible light-infrared-depth is formed; Based on an image-to-image cross-modal registration method, the infrared temperature distribution is aligned with the visible light texture, and early features of hidden cracks caused by wet stains are compensated; Sparse depth maps are generated from binocular image sequences, which are input into a detection model together with visible light images, to realize three-dimensional scale recovery of hidden crack areas.

3. The method for non-destructive testing of defects in a tunnel lining according to claim 1, characterized in that: When performing adaptive denoising on the collected images, first, identify the noise type as Gaussian noise, impulse noise or dust scattering noise, then use the CNN denoising model to perform targeted noise reduction on different noise modes, and calculate the image clarity enhancement factor Ec to dynamically adjust the contrast and brightness of the image, so that the hidden crack edge is prominent, and the accuracy of subsequent crack skeleton extraction is ensured.

4. The method for non-destructive testing of defects in a tunnel lining according to claim 3, characterized in that: Image sharpness enhancement factor E c It is calculated by the following formula: 。 5. The method for non-destructive testing of defects in a tunnel lining according to claim 1, characterized in that: Different sizes of crack and hole features are extracted from the preprocessed images by using a multi-scale CNN, and the hidden crack skeleton point set (xi, yi) is sequentially processed in the crack skeleton mapping process, and the physical length Lc of the hidden crack is obtained by the following formula: ； The crack skeleton points are spatially mapped to obtain the real length of the crack in the physical space.

6. The method for non-destructive testing of defects in a tunnel lining according to claim 1, characterized in that: A deep learning model is constructed to simultaneously complete defect classification, positioning and size estimation, defect classification includes cracks, hollowing and falling, transfer learning is used, a model pre-trained on a large-scale data set of road or building cracks is transferred to the tunnel lining scene, and a small amount of tunnel lining defect data is used for fine tuning, image input, feature extraction, defect recognition and positioning tasks are unified in an end-to-end deep learning model, and the original image is directly input into the network to automatically complete detection and recognition.

7. The method for non-destructive testing of defects in a tunnel lining according to claim 1, characterized in that: Based on the detected defect results, a quantitative health assessment value is generated, and a defect severity index D is calculated s , quantifying the current defect severity, based on the defect severity index D s Using reinforcement learning algorithm to predict future defect development trend, calculate the defect development trend prediction value D p (t+Δt) assigns repair priorities to different defects, visualizes defects in three-dimensional models through BIM and AR technology, and automatically generates assessment reports and repair recommendations.

8. A method of non-destructive testing of defects in a tunnel lining according to claim 7, characterized in that: Defect severity index D s is calculated by the following equation: ； Defect development trend prediction value D p (t+Δt) is calculated by the following equation: 。 9. A method of non-destructive testing of a tunnel lining defect according to claim 8, characterized in that: Based on the defect severity and the prediction result, the maintenance priority, risk level and maintenance suggestion are automatically generated, and the result is combined with the three-dimensional BIM model to show the defect position and severity through augmented reality AR mode; A preset severity threshold Ts and a growth threshold Tg; If the current defect severity index D s is greater than or equal to a preset severity threshold Ts, and the predicted defect severity index D p is greater than or equal to a preset growth threshold Tg in the future time interval Δt, it is determined that the lining region corresponding to the defect is a high-priority maintenance region; otherwise, it is determined that the lining region corresponding to the defect is a low-priority maintenance region.

10. A method of non-destructive testing of a tunnel lining defect according to claim 9, characterized in that: A preset risk threshold Tr and an allowable width threshold Wt form a future trend evaluation strategy: If the predicted defect severity index D p If the current crack width W is greater than or equal to the preset risk threshold Tr in the future time interval Δt, and the current crack width W c greater than or equal to the preset allowable width threshold Wt, the lining area corresponding to the defect is marked as a high risk level, and a mandatory maintenance recommendation is triggered; otherwise, the lining area corresponding to the defect is marked as a general risk level, and a routine inspection recommendation is generated.

Images