A Method and Apparatus for Inverting Shotcrete Thickness Based on Multispectral Images

Abstract

This invention relates to the interdisciplinary field of tunnel engineering quality inspection and multispectral image analysis, and discloses a method and apparatus for shotcrete thickness inversion based on multispectral images, simultaneously achieving high-precision and high-efficiency thickness inversion and void identification. The method includes: S1. Acquiring multispectral images of the shotcrete inspection area surface using a multispectral detection device; S2. Preprocessing the acquired raw multispectral images; S3. Based on the preprocessed multispectral images, extracting the mean multi-band reflectance for thickness inversion and the reflectance variation coefficient for void identification, forming a feature vector; S4. Using the feature vector as input, outputting thickness values ​​and void discrimination values ​​using a trained improved BP neural network, performing thickness inversion based on the thickness values, and determining the location and range of void areas based on the void discrimination values; S5. Outputting the thickness statistics and void defect information of the shotcrete inspection area and visualizing them.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of tunnel engineering quality inspection and multispectral image analysis, and in particular to a method and apparatus for inverting the thickness of shotcrete based on multispectral images. Background Technology

[0002] As the core load-bearing structure of the initial support system for tunnels, the thickness of shotcrete directly determines the load-bearing capacity and impermeability of the support structure. If the thickness is insufficient or there are hollow defects, it is easy to cause safety hazards such as instability of the surrounding rock and water seepage in the tunnel. Therefore, thickness detection is a key link in the quality control of tunnel construction.

[0003] Currently, shotcrete thickness detection technologies are mainly divided into three categories: destructive testing, contact non-destructive testing, and optical testing. However, they all have certain limitations in terms of engineering applicability, detection accuracy, and efficiency, making it difficult to meet the full-section, high-precision, and high-efficiency detection requirements of tunnel engineering. Specifically:

[0004] I. Destructive testing techniques:

[0005] Destructive testing methods, such as core sampling, are the traditional benchmark in the industry. The principle involves drilling holes in the shotcrete surface, extracting core samples, and directly measuring the thickness. While this technique can obtain accurate thickness data at local points, it has the following drawbacks:

[0006] (1) High risk of structural damage: The drilling process will penetrate the shotcrete support layer, destroy the integrity of the initial support, and form a new water leakage channel; at the same time, stress concentration is likely to occur at the drilling site, which may induce crack expansion under the action of surrounding rock pressure, increasing the construction safety risk.

[0007] (2) Low detection efficiency: Only 1-2 points of thickness data can be obtained in a single core drilling. To complete the full cross-section detection of the tunnel, a large number of detection points need to be set up, and a lot of manpower and time need to be invested, resulting in low detection efficiency.

[0008] (3) High cost of subsequent repairs: After inspection, the borehole needs to be grouted and sealed, and the repair quality is difficult to match the original support structure. It may still become a hidden danger in long-term operation.

[0009] II. Contact Non-destructive Testing Technology:

[0010] To replace core drilling, the industry has developed contact non-destructive testing technologies such as spring testing and ultrasonic testing, but both are still limited by operational procedures and environmental adaptability.

[0011] (1) The rebound method indirectly estimates the thickness by measuring the rebound value of the concrete surface. It is affected by the state of the concrete surface. When there is laitance or dust on the surface, the correlation between the rebound value and the thickness is significantly reduced, and the error is large.

[0012] (2) The ultrasonic method requires applying a coupling agent to the concrete surface and calculating the thickness by the propagation time of the sound wave in the concrete. The operation is cumbersome and easily affected by the tunnel environment. In addition, the hollow area will cause multiple reflections of the sound wave, making the calculated thickness value larger than the actual value. It is impossible to distinguish the signal difference between insufficient thickness and hollow area to accurately identify the defect.

[0013] III. Optical Detection Technology:

[0014] Ordinary optical inspection methods can only capture surface morphology and cannot penetrate the concrete surface to obtain internal thickness information, making them extremely unsuitable for dark-colored or rough-surfaced concrete.

[0015] In recent years, multispectral technology has been gradually applied in the field of concrete composition detection due to its advantage of "band-specific identification". Some studies have attempted to use multispectral technology for concrete detection, such as analyzing concrete surface cracks through visible light-near infrared band images. However, in terms of shotcrete thickness inversion, the existing technology is still in its early stages: some schemes only use a single near-infrared band to establish a linear relationship between reflectivity and thickness, without considering the "band-dependent" absorption differences of near-infrared light by concrete of different thicknesses, resulting in large thickness inversion errors; at the same time, existing schemes have not established a correlation identification model between thickness and hollow defects, requiring additional equipment to detect hollow defects, making the detection process still relatively cumbersome. Summary of the Invention

[0016] The technical problem to be solved by the present invention is to provide a method and apparatus for inverting the thickness of shotcrete based on multispectral images, which can simultaneously achieve high-precision and high-efficiency thickness inversion and void identification.

[0017] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0018] On the one hand, the present invention provides a method for inverting the thickness of shotcrete based on multispectral images, comprising the following steps:

[0019] S1. Acquire multispectral images of the surface of the sprayed concrete inspection area using a multispectral detection device;

[0020] S2. Preprocess the acquired raw multispectral images;

[0021] S3. Based on the preprocessed multispectral image, extract the mean multiband reflectance for thickness inversion and the reflectance variation coefficient for hollow area identification to form a feature vector;

[0022] S4. Using the feature vector as input, the improved BP neural network that has been trained outputs thickness value and hollowness discrimination value. Based on the thickness value, thickness inversion is performed, and based on the hollowness discrimination value, the location and range of the hollowness region are determined.

[0023] S5. Output the thickness statistics of the shotcrete inspection area and the information on hollow defects, and display them visually.

[0024] Furthermore, in step S1, the multispectral detection device uses a multispectral camera containing 16 continuous bands with a band spacing of 50nm. When acquiring multispectral images of the surface of the sprayed concrete detection area, the distance between the multispectral camera and the sprayed concrete surface is adjusted to 1.5-2m, and the optical axis of the lens is perpendicular to the concrete surface. Multiple frames of images are continuously acquired according to the preset exposure time, and the average value is taken as the original image data.

[0025] Furthermore, in step S2, the preprocessing of the acquired raw multispectral image includes:

[0026] The original multispectral image was sequentially subjected to noise removal, radiometric correction, and image registration.

[0027] The noise removal employs a Gaussian filtering algorithm, and the formula for calculating the Gaussian filter kernel function is as follows:

[0028] ;

[0029] in, Let be the value of the filter kernel at pixel coordinates (x, y). Let x be the standard deviation, and y be the x and y coordinates of the pixel of the filter kernel, respectively.

[0030] The radiometric correction is performed using a standard white board: first, a multispectral image of the standard white board is acquired, and the standard grayscale values ​​for each band are obtained. Then, acquire multispectral images of the target concrete area to obtain the original grayscale values ​​of each band. The corrected reflectance is calculated using the following formula. :

[0031] ;

[0032] in, For wavelengths in the multispectral band, To correct the concrete at wavelength Reflectance at that location;

[0033] The image registration adopts a SIFT feature point-based registration method: extract SIFT feature points in the visible light band image and record the coordinates of each feature point; match the feature points corresponding to the SIFT feature points in the near-infrared band image; and perform translation and rotation correction on each near-infrared band image through the affine transformation matrix so that the feature point coordinate deviation of each band image is ≤1 pixel.

[0034] Furthermore, in step S3, the extraction of the multi-band reflectivity mean for thickness inversion includes:

[0035] Calculate the average reflectance of each pixel in the detection area across 16 near-infrared bands. As a characteristic of thickness inversion:

[0036] ;

[0037] in, This represents the average multi-band reflectance of a single pixel. The value interval is 50nm, that is, the values ​​are 900nm, 950nm, ... 1650nm, 1700nm respectively;

[0038] The extraction of the reflectance variation coefficient for hollow drum identification includes:

[0039] Calculate the reflectance variation coefficient (CV) of a 3×3 pixel window within the detection area, and use it as a feature for hollow drum detection:

[0040] ;

[0041] in, The mean reflectance of the i-th pixel within a 3×3 window; It is the average of the reflectance values ​​of 9 pixels within a 3×3 window;

[0042] In step S3, the resulting feature vector is represented as: .

[0043] Furthermore, in step S4, the improved BP neural network includes an input layer, a hidden layer, and an output layer; wherein the input layer contains 2 neurons, corresponding to the mean reflectance and the coefficient of variation of reflectance in multiple bands, respectively; the hidden layer contains 12 neurons, and the activation function is the Sigmoid function; the output layer contains 2 neurons, which output the thickness value and the hollowness discrimination value, respectively, wherein the activation function of the neuron outputting the thickness value is a linear function, and the activation function of the neuron outputting the hollowness discrimination value is the Sigmoid function.

[0044] Furthermore, in step S4, the training method of the improved BP neural network includes:

[0045] Multispectral images of shotcrete of different thicknesses were collected, and the measured thickness and void state of each sample were obtained simultaneously through core drilling to obtain a training sample set. The training samples in the training sample set cover different temperature and humidity conditions, and multiple sets of training samples correspond to each thickness of shotcrete.

[0046] Based on the training sample set, the improved BP neural network is trained. During the training process, a momentum term is introduced to update the weights, and the learning rate is dynamically adjusted according to the error changes.

[0047] The introduction of momentum terms for weight updates includes:

[0048] The weight update method from the input layer to the hidden layer is as follows:

[0049] ;

[0050] The weight update method from the hidden layer to the output layer is as follows:

[0051] ;

[0052] in, , , These are the weights from hidden layer neuron i to output layer neuron j during the (k-1)th, kth, and (k+1)th iterations, respectively. , , These are the weights from input layer neuron l to hidden layer neuron i during the (k-1)th, kth, and (k+1)th iterations, respectively. Let be the learning rate for the k-th iteration; For the error term of neuron j in the output layer; This represents the error term for hidden layer neuron i. The output of hidden layer neuron i. This is the input to neuron l in the input layer; The value is 0.9, which is the momentum coefficient;

[0053] The method of dynamically adjusting the learning rate based on error changes includes:

[0054] ;

[0055] in, Let $\frac{k}{k}$ be the network error in the kth iteration. is the learning rate for the (k+1)th iteration.

[0056] Furthermore, the improved training method for the BP neural network also includes:

[0057] During training, mean squared error is used as the objective function for network training, and the calculation formula is as follows:

[0058] ;

[0059] Where N is the number of training samples. This represents the actual thickness value of the t-th sample. The model predicts the thickness value for the t-th sample. Let be the actual void discrimination value of the t-th sample. The model predicts the hollow discriminant value for the t-th sample.

[0060] Further, in step S4, the improved BP neural network, after training, outputs thickness values ​​and void discrimination values. Thickness inversion is performed based on the thickness values, and the location and extent of the void region are determined based on the void discrimination values, including:

[0061] The model outputs the predicted thickness for each pixel. Discrimination value with hollowness Then proceed with the following processing:

[0062] Thickness inversion: for predicted thickness Perform 5×5 pixel window mean filtering to obtain thickness distribution data of the detection area;

[0063] Hollow drum identification: When the hollow drum discrimination value When the value is ≥0.5, the region where the pixel is located is determined to be a hollow area. The area and location of the hollow area are counted and marked as a defect area.

[0064] Furthermore, in step S5, the output of the thickness statistics and hollow defect information of the shotcrete inspection area includes: outputting the average thickness, minimum thickness, and maximum thickness of the inspection area, as well as the number, total area, and maximum hollow size of the hollow areas in the form of a numerical report;

[0065] The visualization includes: displaying a thickness distribution cloud map and a hollow defect marking map. The thickness distribution cloud map uses color gradients to represent different thickness ranges, and the hollow defect marking map uses red boxes to mark hollow areas.

[0066] On the other hand, the present invention also provides a sprayed concrete thickness inversion device based on multispectral images for implementing the above method. The device includes: a multispectral detection device, a data processing terminal, a power supply module, and a retractable gimbal support.

[0067] The multispectral detection device includes a multispectral camera, a near-infrared LED array light source, and a dustproof and waterproof housing;

[0068] The dustproof and waterproof outer shell is wrapped around the outside of the multispectral camera to isolate dust and moisture inside the tunnel;

[0069] The near-infrared LED array light source is matched with the multispectral camera, and the emission wavelength is matched with the camera band, so that a stable reflected light signal can still be obtained when the lighting is uneven in the tunnel.

[0070] The retractable gimbal bracket can rotate 360° horizontally and its height is adjustable. The multispectral detection device is fixed on the bracket. The distance and angle between the multispectral camera and the concrete surface are adjusted by the bracket so that the optical axis of the lens is perpendicular to the concrete surface.

[0071] The power supply module includes a lithium battery pack and a portable solar charging panel, which are electrically connected to the multispectral detection device and the data processing terminal to provide power supply in the tunnel when there is no external power source.

[0072] The data processing terminal is connected to the multispectral detection device via wireless transmission. It has built-in multispectral data preprocessing software and an improved BP neural network model to receive the raw image data transmitted by the multispectral detection device. After preprocessing, it outputs the thickness distribution and hollow defect identification results through model inversion.

[0073] The beneficial effects of this invention are:

[0074] (1) Achieve non-destructive testing and ensure the integrity of the support structure:

[0075] This invention utilizes non-contact multispectral image acquisition, eliminating the need for drilling holes or applying coupling agent to the concrete surface throughout the process. This completely avoids physical damage to the shotcrete support layer, and no repair work is required after the inspection, reducing project costs and construction risks. Furthermore, the non-contact acquisition method ensures that the inspection process does not conflict with the construction schedule, allowing inspections to be completed simultaneously during tunneling intervals without interrupting construction, thus improving project progress efficiency.

[0076] (2) Improve the accuracy of thickness inversion and reduce the impact of environmental interference:

[0077] This invention filters the 900-1700nm near-infrared range (including 16 consecutive bands), covering the 900-1100nm band that is sensitive to 50-100mm thin-layer concrete, and also including the 1100-1700nm band that is suitable for 100-200mm thick-layer concrete. This ensures that the reflectivity differences of concrete of different thicknesses can be accurately captured, avoiding the thickness response blind zone of a single band.

[0078] During the modeling process, a momentum term is introduced to accelerate convergence and avoid local minima. The iteration step size is dynamically adjusted through an adaptive learning rate. By combining the mean reflectance and the coefficient of variation as dual feature inputs, a nonlinear inversion model is constructed. Compared with traditional linear fitting or single feature models, the thickness inversion error is significantly reduced, and it can effectively offset the distortion of reflectance data caused by tunnel dust and uneven illumination. Even under complex working conditions, it can still maintain high accuracy.

[0079] (3) Achieve simultaneous identification of thickness and hollow areas, simplifying the detection process:

[0080] Existing technologies require separate thickness detection equipment and void detection equipment to detect concrete thickness and identify voids, which is cumbersome and costly. This invention, based on the dual feature correlation of multispectral reflectance and an improved BP neural network multi-output design, can simultaneously infer concrete thickness and identify the location and area of ​​voids, significantly reducing equipment investment and labor costs. It is particularly suitable for the need for rapid quality assessment of the entire tunnel cross-section.

[0081] (4) Improve detection efficiency and achieve full cross-section coverage:

[0082] Existing point-by-point inspection methods are extremely inefficient and cannot meet the requirements for full-section inspection of long tunnels. This invention relies on the area array acquisition characteristics of a multispectral camera, which can cover a large area in a single shot, thereby improving inspection efficiency. The multispectral camera used, combined with the rotation and height adjustment functions of the retractable gimbal bracket, can easily cover hard-to-reach areas such as tunnel arches and sidewalls, achieving full-section inspection without blind spots and providing complete data support for the overall assessment of tunnel support quality.

[0083] (5) Enhance environmental adaptability and adapt to complex tunnel working conditions:

[0084] Tunnels often experience uneven lighting, high dust concentrations, and large humidity fluctuations. This invention enhances environmental adaptability through hardware protection and algorithm optimization: At the hardware level, the multispectral camera is equipped with a dustproof and waterproof shell, and the near-infrared LED array light source can adaptively adjust its power according to the light intensity, ensuring that dust and moisture do not affect the cleanliness of the lens and the stability of the light source; at the algorithm level, an image preprocessing workflow based on Gaussian filtering noise reduction, standard whiteboard radiometric correction, and SIFT feature point registration can sequentially eliminate errors caused by dust noise, uneven lighting, and band misalignment, outputting reliable multispectral data suitable for various tunnel engineering scenarios. Attached Figure Description

[0085] Figure 1 This is a schematic diagram of the structure of the sprayed concrete thickness inversion device based on multispectral images in an embodiment of the present invention.

[0086] Figure 2 This is a flowchart of the sprayed concrete thickness inversion method based on multispectral images in an embodiment of the present invention.

[0087] Figure 3 This is a diagram of the improved BP neural network structure in an embodiment of the present invention.

[0088] Figure 1Explanation of markings: 1 is the multispectral camera, 2 is the dustproof and waterproof housing, 3 is the power supply module, 4 is the telescopic gimbal bracket, 5 is the near-infrared LED array light source, 6 is the data processing terminal, and 7 is the sprayed concrete wall surface. Detailed Implementation

[0089] This invention aims to provide a method and apparatus for inverting the thickness of shotcrete based on multispectral images, simultaneously achieving high-precision and high-efficiency thickness inversion and void identification. Its core idea is to construct a technical path based on multispectral band specific response and dual-feature collaborative modeling, encompassing multispectral image acquisition, anti-interference preprocessing, and improved BP neural network multi-output inversion. By correlating the nonlinear relationship between near-infrared reflectivity and thickness, and the characteristic differences between reflectivity fluctuations and voids, high-precision thickness inversion and void defect identification can be achieved simultaneously in a single detection.

[0090] Specifically, the present invention employs the following technical means to achieve the above-mentioned core idea:

[0091] (1) The essential requirement for shotcrete thickness detection is to penetrate the surface to obtain internal thickness information. The absorption and reflection characteristics of near-infrared light of concrete of different thicknesses are band-dependent. For example, 50-100mm thin-layer concrete responds more significantly to 900-1100nm short-wave near-infrared light, while 100-200mm thick-layer concrete is more sensitive to 1100-1700nm long-wave near-infrared light.

[0092] The present invention is based on the thickness sensitivity of multispectral bands, and selects the 900-1700nm near-infrared range (including 16 consecutive bands) to ensure that the thickness variation of concrete can be accurately captured through the reflectivity difference of the corresponding bands within the entire thickness range (50-200mm), laying the physical foundation for subsequent high-precision inversion.

[0093] (2) The multispectral camera array acquisition mode is adopted, which covers a large detection area in a single shot and is non-contact throughout the process. This avoids damage to the support structure and meets the requirements for rapid detection of the entire tunnel section. Since dust, uneven lighting, and humidity fluctuations in the tunnel can cause distortion of multispectral data, a preprocessing process of Gaussian filtering for noise reduction, standard whiteboard radiometric correction, and SIFT feature point registration is designed. Combined with a dustproof and waterproof camera housing and an adaptive LED light source, it is ensured that even in complex environments, the reflectivity data can still accurately reflect the true state of the concrete.

[0094] (3) Based on the dual physical meaning of multispectral reflectance—the mean reflectance reflects the overall absorption of near-infrared light by concrete (which is directly related to the thickness; the greater the thickness, the lower the reflectance), and the reflectance variation coefficient reflects the local reflectance fluctuation (the reflectance fluctuation in the hollow area is much greater than that in the normal area due to the presence of an air layer). By extracting these two types of features to form a feature vector, we can provide a core basis for thickness inversion and a differentiated indicator for hollow area identification.

[0095] (4) In terms of modeling, this invention designs a dual-input-dual-output architecture. The input layer receives the mean reflectance and the coefficient of variation feature vector, and the output layer outputs the thickness value and the hollow discriminant value simultaneously. A complex mapping relationship between reflectance and thickness and hollow is established through nonlinear fitting. At the same time, a momentum term (to accelerate convergence and avoid local minima) and an adaptive learning rate (to dynamically adjust the iteration step size and balance the convergence speed and accuracy) are introduced during the training process to improve the efficiency of model training and the overall accuracy of the model.

[0096] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0097] This embodiment first provides a shotcrete thickness inversion device based on multispectral images, which provides system support for realizing concrete thickness inversion and void identification. See also Figure 1 The device includes a multispectral detection device, a data processing terminal 6, a power supply module 3, and a retractable gimbal bracket 4; wherein, the multispectral detection device includes a multispectral camera 1, a near-infrared LED array light source 5, and a dustproof and waterproof housing 2.

[0098] Multispectral camera 1 employs a high-resolution multispectral camera with an image resolution of 1280×960 pixels, covering the 900-1700nm near-infrared range, comprising 16 consecutive bands (50nm intervals). The 900-1700nm near-infrared band is selected here as an optimized choice based on the traditional near-infrared spectral segmentation (700-1100nm shortwave, 1100-2500nm longwave): it covers the 900-1100nm range in the shortwave, which is sensitive to 50-100mm thin concrete layers, and also includes the 1100-1700nm range in the longwave, which is suitable for 100-200mm thick concrete layers, achieving full thickness range response; at the same time, it eliminates the limited ranges of 700-900nm (shallow penetration, susceptible to visible light interference) and 1700-2500nm (susceptible to water vapor and dust interference), and the mapping relationship between reflectivity and thickness and voids is most significant in this band.

[0099] The multispectral camera 1 is equipped with a dustproof and waterproof outer shell 2 to isolate dust and moisture inside the tunnel and ensure long-term stable image acquisition. The near-infrared LED array light source 5 is matched with the multispectral camera 1, and its emission wavelength is matched with the camera band to obtain stable reflected light signals even when the lighting inside the tunnel is uneven.

[0100] The retractable gimbal bracket 4 can rotate 360° horizontally and is height adjustable. It is used to support the multispectral detection device. The distance and angle between the multispectral camera 1 and the sprayed concrete wall 7 can be adjusted by the bracket so that the optical axis of the lens is perpendicular to the concrete surface. The bracket can be fixed on the tunnel construction platform or the inspection vehicle to realize the acquisition of multi-view images of the entire cross section.

[0101] The power supply module 3 includes a lithium battery pack and a portable solar charging panel, and is electrically connected to the multispectral detection device and the data processing terminal 6 to provide power supply in the absence of an external power source in the tunnel.

[0102] The data processing terminal 6 is an industrial-grade tablet computer that is connected to the multispectral detection device via wireless transmission. It has built-in multispectral data preprocessing software and an improved BP neural network model to receive the raw image data transmitted by the multispectral detection device. After preprocessing, it outputs the thickness distribution and hollow defect identification results through model inversion.

[0103] The device operates as follows: the position and angle of the multispectral camera 1 are adjusted by the telescopic gimbal bracket 4, the near-infrared LED array light source 5 and the multispectral camera 1 are activated, and multispectral images of the sprayed concrete wall 7 are acquired; the image data is transmitted wirelessly to the data processing terminal 6, the terminal first preprocesses the image, then inputs the improved BP neural network model, and finally outputs the thickness distribution cloud map and hollow defect mark. The whole process does not require manual intervention and the detection cycle is short.

[0104] Based on the aforementioned thickness inversion device, this embodiment also provides a method for inverting the thickness of shotcrete based on multispectral images, see [link to relevant documentation]. Figure 2 It includes the following implementation process:

[0105] 1. Multispectral image acquisition:

[0106] In this step, a multispectral detection device is used to acquire multispectral images of the surface of the sprayed concrete detection area.

[0107] In one exemplary implementation, the data collection process is as follows:

[0108] Select the inspection area inside the tunnel. First, clean the surface of the concrete to remove dust and water, ensuring there are no obvious obstructions. Fix the multispectral camera using a telescopic gimbal bracket, adjusting the distance between the camera and the concrete surface to 1.5-2m to ensure the image covers the template area and the lens optical axis is perpendicular to the concrete surface. Adjust the LED light source power according to the light intensity inside the tunnel to ensure stable light intensity. Set the camera acquisition parameters: exposure time 10ms, continuously acquire 3 frames of multispectral images, and take the average value as the raw image data for that area to avoid random noise interference from a single acquisition.

[0109] 2. Image preprocessing:

[0110] In this step, the acquired raw multispectral images are preprocessed.

[0111] In one exemplary implementation, since the original multispectral image is susceptible to dust and uneven illumination within the tunnel, this step performs preprocessing through noise removal, radiometric correction, and image registration to ensure the accuracy of the reflectance data.

[0112] (1) Noise Removal:

[0113] Gaussian filtering is used to eliminate high-frequency noise (such as bright spots caused by dust particles) in images. The formula for calculating the Gaussian filter kernel function is as follows:

[0114] ;

[0115] In the formula, The value of the filter kernel at pixel coordinates (x, y); The standard deviation is denoted by 1.5 (balancing noise removal and detail preservation); x and y are the pixel coordinates of the filter kernel. Convolving this filter kernel with the original multispectral image yields the denoised image, reducing noise intensity.

[0116] (2) Radiation correction:

[0117] To eliminate reflectivity deviations caused by uneven lighting, a standard white board is used for radiation correction.

[0118] First, a multispectral image of a standard white board was acquired, and the standard grayscale values ​​of each band were obtained. Then, acquire multispectral images of the target concrete area to obtain the original grayscale values ​​of each band. The corrected reflectance is calculated using the following formula. :

[0119] ;

[0120] In the formula, The wavelengths are for the multispectral bands (range 900-1700nm, interval 50nm). To correct the concrete at wavelength The reflectance at that location (range 0-1).

[0121] (3) Image registration:

[0122] Because the imaging chips of different bands in a multispectral camera have slight displacements, it is necessary to align the images of each band to the same coordinate system. This embodiment adopts a registration method based on feature points:

[0123] SIFT feature points are extracted from visible light images, and the coordinates of each feature point are calculated. Corresponding feature points are then located in other near-infrared images, and translation and rotation corrections are applied to each image using an affine transformation matrix to ensure that the coordinate deviation of feature points in each image is ≤1 pixel. The registered images ensure that reflectivity data from different bands correspond to the same concrete area, avoiding thickness calculation errors caused by band misalignment.

[0124] 3. Feature extraction:

[0125] In this step, based on the preprocessed multispectral image, the mean reflectance of multiple bands for thickness inversion and the coefficient of variation of reflectance for hollow area identification are extracted to form a feature vector.

[0126] In one exemplary implementation, two types of core features are extracted from the preprocessed multispectral image, which are used for thickness inversion and hollow area identification, respectively:

[0127] (1) Thickness-related features:

[0128] Average reflectance across multiple wavelengths: Since concrete of different thicknesses absorbs near-infrared light to varying degrees, the greater the thickness, the shallower the penetration depth of near-infrared light and the lower the reflectance. This trend is most pronounced in the 900-1700nm wavelength range.

[0129] Calculate the average reflectance of each pixel in the detection area across 16 near-infrared bands. As the core feature of thickness inversion, the calculation formula is:

[0130] ;

[0131] In the formula, This represents the average reflectance of a single pixel across multiple bands (range 0-1). The value interval is 50nm (i.e., 900nm, 950nm, ..., 1700nm).

[0132] (2) Hollow-out correlation characteristics:

[0133] Coefficient of variation of reflectance: The concrete in the hollow area is separated from the surrounding rock and contains an air layer. The reflectance of air to near-infrared light is much higher than that of concrete, which causes the reflectance of the hollow area to fluctuate significantly more than that of the normal area.

[0134] The reflectance variation coefficient (CV) of a 3×3 pixel window within the detection area is calculated as a feature for hollow drum identification. The calculation formula is as follows:

[0135] ;

[0136] In the formula, Let be the average reflectance of the i-th pixel within a 3×3 window. is the average reflectance of 9 pixels within a 3×3 window, and CV is the reflectance variation coefficient, which is dimensionless.

[0137] Through the feature extraction described above, a set of feature vectors can be obtained for each pixel. This provides data support for subsequent neural network inputs.

[0138] 4. Improve the construction and training of BP neural networks:

[0139] In this step, a BP neural network model is constructed and trained to obtain a model that can be applied to real-world scenarios to simultaneously perform thickness inversion and hollowness identification.

[0140] In one exemplary implementation, the structure of the improved BP neural network in this embodiment is shown below. Figure 3 It includes an input layer, a hidden layer, and an output layer:

[0141] Input layer: Contains 2 neurons, corresponding to feature vectors Enter the mean reflectance and the coefficient of variation of reflectance respectively.

[0142] Hidden layer: The number of neurons was determined to be 12 using cross-validation, and the sigmoid function was used as the activation function. This enhances the model's nonlinear fitting ability.

[0143] Output layer: Contains 2 neurons, which output "thickness value d" (unit: mm) and "hollow discrimination value k" (k=1 indicates hollowness, k=0 indicates no hollowness). The activation function of the thickness output neuron is a linear function, and the activation function of the hollowness output neuron is a sigmoid function (output value ≥0.5 is judged as hollowness).

[0144] In addition, to avoid the problems of slow convergence speed and easy getting trapped in local minima, this embodiment optimizes the network convergence speed and generalization ability by introducing a momentum term and an adaptive learning rate.

[0145] Momentum term introduction: The change in weights from the previous iteration is incorporated into the weight update, accelerating convergence and avoiding local minima.

[0146] The weight update method from the input layer to the hidden layer is as follows:

[0147] ;

[0148] The weight update method from the hidden layer to the output layer is as follows:

[0149] ;

[0150] in, , , These are the weights from hidden layer neuron i to output layer neuron j during the (k-1)th, kth, and (k+1)th iterations, respectively. , , These are the weights from input layer neuron l to hidden layer neuron i during the (k-1)th, kth, and (k+1)th iterations, respectively. Let be the learning rate for the k-th iteration; For the error term of neuron j in the output layer; This represents the error term for hidden layer neuron i. The output of hidden layer neuron i. This is the input to neuron l in the input layer; The value is 0.9, which is the momentum coefficient.

[0151] Adaptive learning rate adjustment: The learning rate is dynamically adjusted based on error changes to avoid oscillations caused by an excessively large learning rate or slow convergence caused by an excessively small learning rate. The adjustment formula is as follows:

[0152] ;

[0153] In the formula, Let be the total network error in the k-th iteration, and be the initial learning rate. .

[0154] Error function definition: The mean squared error is used as the objective function for network training, and the calculation formula is as follows:

[0155] ;

[0156] In the formula, N is the number of training samples. This represents the actual thickness value of the t-th sample (obtained through core drilling, with an accuracy of ±0.1 mm). The model predicts the thickness value for the t-th sample. This is the actual hollow discriminant value for the t-th sample (1 for hollow, 0 for no hollow). The model predicts the hollow discriminant value for the t-th sample.

[0157] Based on the improved BP neural network construction described above, training samples were constructed using measured data from tunnel engineering projects:

[0158] Collect multispectral images of shotcrete with different thicknesses ranging from 50 to 150 mm (intervals of 10 mm), core-drilled measured thickness d, and void status k (1 for void, 0 for no void), covering different temperature and humidity conditions. Each thickness contains dozens of data sets (each set contains a certain number of voids). Extract feature vectors X to construct a sample set.

[0159] The sample set is divided into a training set and a validation set according to a certain ratio, and the network weights are initialized. , Use random numbers in the interval [-0.1, 0.1]; substitute them into the training set samples for iterative training, and calculate the error on the validation set every 100 iterations. ,when Training is stopped and network parameters are saved when the network stops decreasing after 5 consecutive iterations or when the number of iterations reaches 5000.

[0160] 5. Thickness inversion and hollowness identification:

[0161] In this step, based on the improved BP neural network that has been trained, thickness inversion and hollow area identification are performed on the region to be detected.

[0162] In one exemplary implementation, the multispectral image feature vector of the detection region is preprocessed. Input the trained improved BP neural network, and the model outputs the predicted thickness for each pixel. Discrimination value with hollowness .

[0163] Thickness inversion: for predicted thickness Smoothing is performed (5×5 pixel window mean filtering) to eliminate errors from isolated pixels, and the thickness distribution data of the detection area is obtained. The thickness value range needs to cover the design thickness range of the tunnel shotcrete.

[0164] Hollow drum identification: When the hollow drum discrimination value When the pixel is located, the region is determined to be hollow, the area and location of the hollow region are counted, and it is marked as a defect region.

[0165] 6. Output Results:

[0166] In this step, the thickness statistics of the shotcrete inspection area and the information on hollow defects are output and visualized.

[0167] In one exemplary implementation, the detection results are output in two forms on the data processing terminal:

[0168] Numerical Report: Outputs the average thickness, minimum thickness, and maximum thickness of the detection area, as well as the number of hollow areas, total area, and maximum hollow size.

[0169] Image visualization: Generate thickness distribution cloud map and hollow defect marking map, use color gradient to represent thickness, and use red boxes to mark hollow areas.

[0170] Although embodiments of the present invention have been described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present invention, and all such changes and alterations shall not depart from the protection scope of the present invention.

Claims

1. A method for inverting the thickness of shotcrete based on multispectral images, characterized in that, Includes the following steps: S1. Acquire multispectral images of the surface of the sprayed concrete inspection area using a multispectral detection device; S2. Preprocess the acquired raw multispectral images; S3. Based on the preprocessed multispectral image, extract the mean multiband reflectance for thickness inversion and the reflectance variation coefficient for hollow area identification to form a feature vector; The extraction of the multi-band reflectivity mean for thickness inversion includes: Calculate the average reflectance of each pixel in the detection area across 16 near-infrared bands. As a characteristic of thickness inversion: ； in, This represents the average multi-band reflectance of a single pixel. For wavelengths in the multispectral band, The value interval is 50nm, that is, the values ​​are 900nm, 950nm, ... 1650nm, 1700nm respectively; To correct the concrete at wavelength Reflectance at that location; The extraction of the reflectance variation coefficient for hollow drum identification includes: Calculate the reflectance variation coefficient (CV) of a 3×3 pixel window within the detection area, and use it as a feature for hollow drum detection: ； in, The mean reflectance of the i-th pixel within a 3×3 window; It is the average of the reflectance values ​​of 9 pixels within a 3×3 window; In step S3, the resulting feature vector is represented as: ; S4. Using the feature vector as input, the improved BP neural network that has been trained outputs thickness value and hollowness discrimination value. Based on the thickness value, thickness inversion is performed, and based on the hollowness discrimination value, the location and range of the hollowness region are determined. The improved BP neural network includes an input layer, a hidden layer, and an output layer. The input layer contains two neurons, corresponding to the mean reflectance and the coefficient of variation of reflectance across multiple bands, respectively. The hidden layer contains twelve neurons, with the sigmoid function as the activation function. The output layer contains two neurons, outputting a thickness value and a hollowness discrimination value, respectively. The activation function of the neuron outputting the thickness value is a linear function, while the activation function of the neuron outputting the hollowness discrimination value is a sigmoid function. S5. Output the thickness statistics of the shotcrete inspection area and the information on hollow defects, and display them visually.

2. The method for inverting shotcrete thickness based on multispectral images as described in claim 1, characterized in that, In step S1, the multispectral detection device uses a multispectral camera containing 16 continuous bands with a band spacing of 50nm. When acquiring multispectral images of the surface of the sprayed concrete detection area, the distance between the multispectral camera and the sprayed concrete surface is adjusted to 1.5-2m, and the optical axis of the lens is perpendicular to the concrete surface. Multiple frames of images are continuously acquired according to the preset exposure time, and the average value is taken as the original image data.

3. The method for inverting shotcrete thickness based on multispectral images as described in claim 1, characterized in that, In step S2, the preprocessing of the acquired raw multispectral image includes: The original multispectral image was sequentially subjected to noise removal, radiometric correction, and image registration. The noise removal employs a Gaussian filtering algorithm, and the formula for calculating the Gaussian filter kernel function is as follows: ； in, Let be the value of the filter kernel at pixel coordinates (x, y). Let x be the standard deviation, and y be the x and y coordinates of the pixel of the filter kernel, respectively. The radiometric correction is performed using a standard white board: first, a multispectral image of the standard white board is acquired, and the standard grayscale values ​​for each band are obtained. Then, acquire multispectral images of the target concrete area to obtain the original grayscale values ​​of each band. The corrected reflectance is calculated using the following formula. : ; The image registration adopts a SIFT feature point-based registration method: extract SIFT feature points in the visible light band image and record the coordinates of each feature point; match the feature points corresponding to the SIFT feature points in the near-infrared band image; and perform translation and rotation correction on each near-infrared band image through the affine transformation matrix so that the feature point coordinate deviation of each band image is ≤1 pixel.

4. The method for inverting shotcrete thickness based on multispectral images as described in claim 1, characterized in that, In step S4, the training method of the improved BP neural network includes: Multispectral images of shotcrete of different thicknesses were collected, and the measured thickness and void state of each sample were obtained simultaneously through core drilling to obtain a training sample set. The training samples in the training sample set cover different temperature and humidity conditions, and multiple sets of training samples correspond to each thickness of shotcrete. Based on the training sample set, the improved BP neural network is trained. During the training process, a momentum term is introduced to update the weights, and the learning rate is dynamically adjusted according to the error changes. The introduction of momentum terms for weight updates includes: The weight update method from the input layer to the hidden layer is as follows: ； The weight update method from the hidden layer to the output layer is as follows: ； in, , , These are the weights from hidden layer neuron i to output layer neuron j during the (k-1)th, kth, and (k+1)th iterations, respectively. , , These are the weights from input layer neuron l to hidden layer neuron i during the (k-1)th, kth, and (k+1)th iterations, respectively. Let be the learning rate for the k-th iteration; This represents the error term for neuron j in the output layer; This represents the error term for hidden layer neuron i. The output of hidden layer neuron i. This is the input to neuron l in the input layer; The value is 0.9, which is the momentum coefficient; The method of dynamically adjusting the learning rate based on error changes includes: ； in, Let $\mathbf{k}$ be the network error in the kth iteration. is the learning rate for the (k+1)th iteration.

5. The method for inverting shotcrete thickness based on multispectral images as described in claim 4, characterized in that, The improved training method for the BP neural network also includes: During training, mean squared error is used as the objective function for network training, and the calculation formula is as follows: ； Where N is the number of training samples. This represents the actual thickness value of the t-th sample. The model predicts the thickness value for the t-th sample. Let be the actual void discrimination value of the t-th sample. The model predicts the hollow discriminant value for the t-th sample.

6. The method for inverting shotcrete thickness based on multispectral images as described in claim 5, characterized in that, In step S4, the improved BP neural network, after training, outputs thickness values ​​and void discrimination values. Thickness inversion is performed based on the thickness values, and the location and extent of the void region are determined based on the void discrimination values. This includes: The model outputs the predicted thickness for each pixel. Discrimination value with hollowness Then proceed with the following processing: Thickness inversion: for predicted thickness Perform 5×5 pixel window mean filtering to obtain thickness distribution data of the detection area; Hollow drum identification: When the hollow drum discrimination value When the value is ≥0.5, the region where the pixel is located is determined to be a hollow area. The area and location of the hollow area are counted and marked as a defect area.

7. The method for inverting shotcrete thickness based on multispectral images as described in any one of claims 1-6, characterized in that, In step S5, the output of the thickness statistics and hollow defect information of the shotcrete inspection area includes: outputting the average thickness, minimum thickness, and maximum thickness of the inspection area, as well as the number, total area, and maximum hollow size of the hollow areas in the form of a numerical report; The visualization includes: displaying a thickness distribution cloud map and a hollow defect marking map. The thickness distribution cloud map uses color gradients to represent different thickness ranges, and the hollow defect marking map uses red boxes to mark hollow areas.

8. A shotcrete thickness inversion device based on multispectral images, used to implement the shotcrete thickness inversion method based on multispectral images as described in any one of claims 1-7, characterized in that, The device includes: a multispectral detection device, a data processing terminal, a power supply module, and a retractable gimbal support. The multispectral detection device includes a multispectral camera, a near-infrared LED array light source, and a dustproof and waterproof housing; The dustproof and waterproof outer shell is wrapped around the outside of the multispectral camera to isolate dust and moisture inside the tunnel; The near-infrared LED array light source is matched with the multispectral camera, and the emission wavelength is matched with the camera band, so that a stable reflected light signal can still be obtained when the lighting is uneven in the tunnel. The retractable gimbal bracket can rotate 360° horizontally and its height is adjustable. The multispectral detection device is fixed on the bracket. The distance and angle between the multispectral camera and the concrete surface are adjusted by the bracket so that the optical axis of the lens is perpendicular to the concrete surface. The power supply module includes a lithium battery pack and a portable solar charging panel, which are electrically connected to the multispectral detection device and the data processing terminal to provide power supply in the tunnel when there is no external power source. The data processing terminal is connected to the multispectral detection device via wireless transmission. It has built-in multispectral data preprocessing software and an improved BP neural network model to receive the raw image data transmitted by the multispectral detection device. After preprocessing, it outputs the thickness distribution and hollow defect identification results through model inversion.

Images