Tunnel segment near-surface defect detection method based on hybrid deep learning network

Abstract

The invention provides a tunnel segment near-surface defect detection method based on a hybrid deep learning network, and relates to the field of tunnel engineering detection.The method comprises the steps that S1, an impact echo method is used for collecting tunnel segment test block near-surface defect burial depth signals; s2, inputting the defect burial depth signal into a first preset neural network and a second preset neural network to obtain a first predicted burial depth and a second predicted burial depth; s3, calculating an error average value of the defect burial depth signal, the shallow burial range signal and the deep burial range signal according to an average value of difference values between the first predicted burial depth and the actual burial depth and between the second predicted burial depth and the actual burial depth; and S4, calculating the weight ratio of the defect burial depth signal, the shallow burial range signal and the deep burial range signal in the two neural networks, and then calculating and outputting the final burial depth according to an adaptive multi-scale weighting algorithm. According to the method, the burial depth of the defect is calculated through the hybrid neural network, the near-surface defect of the tunnel segment is detected, and the accuracy and efficiency of tunnel segment detection are improved.

Description

Technical Field

[0001] The present invention relates to the field of tunnel engineering detection, and in particular to a tunnel segment near-surface defect detection method based on a hybrid deep learning network. Background Art

[0002] In recent years, with the continuous development of road traffic and the complexity of geological conditions in my country, tunnel engineering has been widely used to solve the needs of road construction in complex geological environments. The shield segment is the main assembly component of the tunnel shield construction and the innermost barrier of the tunnel. It bears the role of resisting soil pressure, groundwater pressure and some special loads, and has a direct and important impact on the structural performance, durability and waterproof performance of the tunnel. Reference 1. However, during the manufacturing and construction process of the segment, defects and quality problems are inevitable, such as cracks, voids, roughness and uneven thickness, which will not only affect the bearing capacity and structural stability of the tunnel segment, but also may cause leakage, corrosion and damage, and thus cause unpredictable effects on the operation safety and life of the entire tunnel project. Reference 2. Therefore, it is very important to detect and evaluate the potential defects inside the tunnel segment in a timely manner.

[0003] The impact echo method (IE) was developed in the late 1980s and is a very effective means of detecting internal defects in structures. IE instruments began to be marketed in the late 1990s. The American Society for Testing and Materials (ASTM) adopted this technology in 1998 and officially used it to measure the thickness of concrete structures. Zhang Jingbin of Beijing Jiaotong University analyzed the characteristic spectrum of different concrete specimens based on the impact echo method and established a basic calculation formula for concrete wave velocity and strength, which can be used to deduce the compressive strength value of concrete structures, Reference 3. Dorafshan and Azari used the impact echo method to detect reinforced concrete bridge samples with artificial defects, laboratory-made concrete samples with artificial underground defects, and covering systems made of cement and asphalt covering materials, and achieved relatively accurate detection results, providing an effective method for detecting the quality of concrete structures, References 4 and 5.

[0004] With the continuous development of Internet technology, various computer parameters such as computing power and computing speed have been gradually upgraded, and the large-scale computing needs of deep learning have been met. More and more algorithm models have been applied to damage identification and detection, Reference 6. Deep learning networks do not need to manually set feature vectors based on image features, and can be used for the detection and identification of complex tunnel damage. Zhu Hongchen et al. trained and tested the constructed crack data set based on the Faster-RCNN target detection algorithm. The results showed that the algorithm is applicable to the lining crack detection task of actual tunnel engineering, Reference 7; Kumar et al. used the YOLO-v3 deep learning model for real-time detection of concrete damage in high-rise civil structures. The proposed high-rise structure real-time damage detection system based on drones can be used to test concrete cracks of different shapes and sizes and the spalling of high-rise concrete structures, Reference 8; Wu Hehe et al. proposed a tunnel crack detection method based on Faster R-CNN, which verified the possibility and accuracy of tunnel crack detection using deep learning methods, Reference 9.

[0005] The shield segment is the permanent lining structure of the shield tunnel. The quality of the shield segment is directly related to the overall quality and safety of the tunnel, and affects the waterproof performance and durability of the tunnel. In engineering practice, when there are defects near the surface of the tunnel segment, the existing detection method can only detect the type of defect but cannot locate the defect well.

[0006] References:

[0007] [1] Wang Zhenxin. Durability of shield tunnels[J]. Underground Engineering and Tunneling, 2002(02):2-5+49.

[0008] [2] Ye Zhi, Fu Anran, Liu Huabei. Influence of seepage erosion of shield tunnel vault on surface settlement and structural deformation[J]. Journal of Hohai University (Natural Science Edition), 2021, 49(03): 279-287.

[0009] [3] Zhang Jingbin. Research on application of impact echo method in nondestructive testing of prestressed concrete structures[D]. Beijing Jiaotong University, 2018.

[0010] [4] Dorafshan S, Azari H. Deep learning models for bridge deck evaluation using impact echo[J]. Construction and Building Materials, 2020, 263: 120109.

[0011] [5] Dorafshan S, Azari H. Evaluation of bridge decks with overlays using impact echo, a deep learning approach [J]. Automation in Construction, 2020, 113: 103133.

[0012] [6] Xue Yadong, Li Yicheng. Shield tunnel lining defect identification method based on deep learning[J]. Journal of Hunan University (Natural Science Edition), 2018, 45(03): 100-109. DOI: 10.16339 / j.cnki.hdxbzkb.2018.03.012.

[0013] [7] Zhu Hongchen, Liu Yuchu. Intelligent identification of tunnel lining cracks based on deep learning[J]. Smart City, 2024, 10(03): 12-14. DOI: 10.19301 / j.cnki.zncs.2024.03.004.

[0014] [8] P. Kumar, S. Batchu, N. Swamy S. and SR Kota, "Real-Time Concrete DamageDetection Using Deep Learning for High Rise Structures," in IEEE Access, vol.9, pp.112312-112331, 2021, doi:10.1109 / ACCESS.2021.3102647.

[0015] [9] Wu Hehe, Wang Anhong, Wang Haidong. Tunnel image crack detection based on Faster R-CNN[J]. Journal of Taiyuan University of Science and Technology, 2019, 40(03): 165-168.

[0016] The background description provided herein is for the purpose of generally presenting the context of the present disclosure. Unless otherwise indicated herein, the materials described in this section are not prior art to the claims of this application and are not admitted to be prior art by inclusion in this section. Summary of the invention

[0017] In order to overcome the shortcomings in the background technology, the present invention discloses a tunnel segment near-surface defect detection method based on a hybrid deep learning network.

[0018] In order to achieve the above-mentioned invention object, the present invention adopts the following technical scheme:

[0019] The tunnel segment near-surface defect detection method based on hybrid deep learning network includes the following steps:

[0020] S1. Use the impact echo method to collect the buried depth signal of the near-surface defects of the tunnel segment test block;

[0021] S2, inputting the defect burial depth signal into the first preset neural network and the second preset neural network to obtain the first predicted burial depth and the second predicted burial depth; the defect burial depth signal includes a shallow burial range signal and a deep burial range signal; the first predicted burial depth includes a first shallow burial predicted burial depth and a first deep burial predicted burial depth; the second predicted burial depth includes a second shallow burial predicted burial depth and a second deep burial predicted burial depth; the first preset neural network and the second preset neural network are different types of neural networks;

[0022] S3, calculating the average error of the defect burial depth signal, the shallow burial range signal and the deep burial range signal in the first preset neural network and the second preset neural network according to the average value of the difference between the first predicted burial depth and the second predicted burial depth and the actual burial depth;

[0023] S4. Reciprocally normalize the error averages of the defect depth signal, shallow burial range signal and deep burial range signal in the first preset neural network and the second preset neural network to obtain their weight ratios in the two neural networks, and then calculate and output the final burial depth according to the adaptive multi-scale weighted algorithm.

[0024] Specifically, step S3 specifically includes the following steps:

[0025] S31, respectively calculating the difference between the first predicted burial depth y1 and the second predicted burial depth y2 and the actual burial depth, and then averaging the differences to obtain a first error average value and the second error mean

[0026] S32, respectively calculate the difference between the first shallow buried predicted burial depth y_l1 and the second shallow buried predicted burial depth y_l2 and the actual burial depth, and then average the differences to obtain the first shallow buried error average value and the second shallow buried error average

[0027] S33, respectively calculating the difference between the first deep burial predicted depth y_h1 and the second deep burial predicted depth y_h2 and the actual burial depth, and then averaging the differences to obtain the first deep burial error average value and the average value of the second deep buried error

[0028] Specifically, step S4 specifically includes the following steps:

[0029] S41, reciprocally normalizing the error averages of the defect depth signal in the first preset neural network and the second preset neural network to obtain a first depth weight ratio Q1 and a second depth weight ratio Q2;

[0030] S42, reciprocally normalizing the error averages of the shallow burial range signal in the first preset neural network and the second preset neural network to obtain a first shallow burial weight ratio Q_l1 and a second shallow burial weight ratio Q_l2;

[0031] S43, reciprocally normalizing the error averages of the deep burial range signal in the first preset neural network and the second preset neural network to obtain a first deep burial weight ratio Q_h1 and a second deep burial weight ratio Q_h2;

[0032] S44, calculating and outputting the final burial depth according to an adaptive multi-scale weighted algorithm; the adaptive multi-scale weighted algorithm is as follows:

[0033]

[0034] Where y is the final burial depth; thr is the preset burial depth threshold, thr = (d max -d min )÷2+d min , d min is the minimum burial depth; d max is the maximum burial depth.

[0035] Specifically, step S1 is as follows: impact echo detection is performed near the surface of the tunnel segment test block, and then acoustic wave data is collected by an air-coupled probe to obtain a defect burial depth signal.

[0036] Specifically, in step S1, defects of the same type and different buried depths are buried in the tunnel segment test block; the defects are filled with hard foam materials.

[0037] Specifically, in step S2, the first preset neural network obtains the corresponding first predicted burial depth according to the input defect burial depth signal; the first preset neural network is trained by obtaining the first time-frequency image data according to the short-time Fourier transform of the defect burial depth signal.

[0038] Specifically, in step S2, the first preset neural network is a ResNeXt model.

[0039] Specifically, in step S2, the second preset neural network obtains the corresponding second predicted burial depth according to the input defect burial depth signal; the second preset neural network is trained by obtaining the second time-frequency image data according to the S transform of the defect burial depth signal.

[0040] Specifically, in step S2, the second preset neural network is a ConvNeXt V2 model.

[0041] Specifically, before the first preset neural network and the second preset neural network are trained, the method further includes: converting the first time-frequency image data and the second time-frequency image data into a tensor format and performing standardization processing through formula (1), which is as follows:

[0042]

[0043] Among them, normalizedpixel is the pixel value after normalization of the image data; pixel is the original pixel value in the image data; mean is the mean vector; std is the standard deviation vector.

[0044] The present invention provides a method for detecting near-surface defects of tunnel segments based on a hybrid deep learning network, comprising the following steps: S1, using an impact echo method to collect near-surface defect burial depth signals of tunnel segment test blocks; S2, inputting the defect burial depth signals into a first preset neural network and a second preset neural network to obtain a first predicted burial depth and a second predicted burial depth; S3, calculating the error average values ​​of the defect burial depth signals, shallow buried range signals and deep buried range signals in the first preset neural network and the second preset neural network according to the average values ​​of the differences between the first predicted burial depth and the second predicted burial depth and the actual burial depth; S4, respectively reciprocally normalizing the error average values ​​of the defect burial depth signals, shallow buried range signals and deep buried range signals in the first preset neural network and the second preset neural network to obtain their weight ratios in the two neural networks, and then calculating and outputting the final burial depth according to an adaptive multi-scale weighted algorithm. The present invention mainly focuses on using an impact echo method based on an air-coupled ultrasonic array to generate and collect acoustic wave data, converting the acoustic wave data into image data through a data processing process, and using a hybrid neural network to calculate the burial depth of the defects, thereby realizing the detection of near-surface defects of tunnel segments and improving the accuracy and efficiency of tunnel segment detection.

[0045] In addition, the average absolute value of the difference between the calculation result of the data collected by the present invention on the hybrid depth prediction model and the actual burial depth is 1.76 mm, which can well locate the position of the near-surface defects of the tunnel segment and has guiding significance for actual engineering practice;

[0046] In addition, the present invention performs a series of standardized preprocessing steps on the tunnel segment defect data images before model training to ensure that each image received by the model has the same scale and distribution. Such processing helps the model to learn and generalize better, helps the model training to be smoother, and makes the model easier to converge.

[0047] In addition, the present invention provides a method for making tunnel segment test blocks, which can simulate the actual situation of tunnel segments under controlled laboratory conditions and allow researchers to accurately control the experimental conditions, making the data more controllable, and the production of laboratory test blocks can follow a standardized process, making the experimental results more comparable. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0049] Figure 1 It is a schematic flow chart of a tunnel segment near-surface defect detection method based on a hybrid deep learning network according to an embodiment of the present invention;

[0050] Figure 2 is a schematic diagram of the principle of the impact echo method provided according to an embodiment of the present invention;

[0051] Figure 3 is a schematic diagram of the structure of an air-coupled sensor array provided according to an embodiment of the present invention;

[0052] Figure 4 is a schematic diagram of a test block design provided according to an embodiment of the present invention;

[0053] Figure 5 It is a schematic diagram of the basic framework of the test block and defect fixing according to an embodiment of the present invention;

[0054] Figure 6 is a schematic diagram of a test block that has been cast according to an embodiment of the present invention;

[0055] Figure 7 is a structural diagram based on a hybrid deep learning network provided according to an embodiment of the present invention;

[0056] Figure 8 is a schematic diagram of STFT transformation provided according to an embodiment of the present invention;

[0057] Fig. 9 is a schematic diagram of the presentation form of the ResNext block provided according to an embodiment of the present invention; wherein Fig. 9 (a) in its original form; Fig. 9 (b) is in Concat form; Fig. 9 (c) is the group convolution form;

[0058] Fig.10 is a schematic diagram of an S transformation provided according to an embodiment of the present invention;

[0059] Fig.11 It is a schematic diagram of a ResNeXt burial depth prediction loss change curve provided according to an embodiment of the present invention. DETAILED DESCRIPTION

[0060] The present invention can be explained in detail through the following embodiments. The purpose of disclosing the present invention is to protect all technical improvements within the scope of the present invention. In the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "front", "back", "left", "right", etc. indicating directions or positional relationships, they only correspond to the drawings of the present application for the convenience of describing the present invention, and do not indicate or imply that the device or element referred to must have a specific direction.

[0061] Embodiment 1

[0062] refer to Figure 1 This embodiment provides a tunnel segment near-surface defect detection method based on a hybrid deep learning network, which specifically includes the following steps:

[0063] S1. Use the impact echo method to collect the buried depth signal of the near-surface defects of the tunnel segment test block;

[0064] refer to Figure 2 The impact echo method is a sophisticated nondestructive testing technology that cleverly uses the propagation characteristics of elastic waves excited by transient impact in the medium to generate acoustic wave data. A typical IE instrument includes an impact source, a nearby receiver, and a data acquisition system. A hammer or other impact device is used to excite instantaneous shock waves on the surface of the structure. These waves propagate inside the concrete and reflect when they encounter cracks, voids, or other defects. These reflected waves are captured by the acoustic wave sensors installed on the surface of the structure, and the sensors transmit the captured signals to the data acquisition card. Subsequently, the data acquisition system receives these signals and uses the data analysis and processing system to convert the time domain waveform of the reflected wave into a frequency domain signal.

[0065] Specifically, impact echo detection is carried out on the surface of the tunnel segment test block and collected by air-coupled probe;

[0066] This embodiment uses an air-coupled probe to collect signals during the impact echo method data collection stage of the tunnel segment test block. The air-coupled probe has non-contact and high sensitivity, which can effectively detect and locate defect areas and avoid the use of coupling agents between the probe and the material to be inspected. The air-coupled sensor array as a receiving device can also improve the detection speed and detection efficiency in engineering applications. Figure 3In the detection of tunnel segments, the data generated by impact echo detection on the surface of the segment can accurately locate the location of the cracks and assess their depth and severity. This method can not only provide an intuitive image of the cracks, but also reveal the morphology and nature of the cracks by analyzing the time-frequency characteristics of the reflected waves. The application of this technology can greatly improve the efficiency and accuracy of tunnel segment detection and provide a scientific basis for the maintenance and repair of tunnel projects.

[0067] Specifically, the tunnel segment test block contains defects of the same type and different buried depths;

[0068] This embodiment also provides a method for making a tunnel segment test block. This embodiment simulates the tunnel segment itself by casting a reinforced concrete test block, and buries hard foam of different depths in the reinforced concrete test block to simulate the "hollow" defect. During the experiment, the test block size designed in this article is 1500mm*1200mm*200mm. The test block defects are filled with hard foam materials, and the size is 100mm×100mm. The filling depth is controlled as a single variable. The hard foam is buried in the range of 10mm deep to 170mm deep, and a filling depth is set every 5mm. The design diagram is as shown in the figure. Figure 4 shown.

[0069] When making reinforced concrete test blocks simulating tunnel segments, steel bars are used to weld basic frames of different depths to fix the position of the landfill material, especially to ensure the accuracy of the defect landfill depth, such as Figure 5 The prepared landfill material is fixed at the set position, and finally concrete is poured on the steel skeleton to form a block, and the test block is completed, as shown in 6. After it is completely solidified, it is used for the test block knocking test.

[0070] The impact echo method based on the air-coupled ultrasonic array is used to collect data. The audio signal collection software is prepared on the computer, and multiple air-coupled sound receiving units are connected to form a specific geometric array. The equipment is fixed near the sound emission position, and the test block is knocked at a certain frequency using an automatic knocking device to generate sound wave data. Since the tunnel segment test block contains defects of different burial depths, the sound wave data contains the burial depth signal of the near-surface defects of the tunnel segment test block. After being collected by the air-coupled sound receiving unit, the computer displays the sound wave waveform for real-time observation of the data validity.

[0071] Since the designed defect positions are marked before signal collection, the actual burial depth corresponding to the collected defect burial depth signal can be known when collecting the signal.

[0072] S2, inputting the defect burial depth signal into the first preset neural network and the second preset neural network to obtain the first predicted burial depth and the second predicted burial depth; the defect burial depth signal includes a shallow burial range signal and a deep burial range signal; the first predicted burial depth includes a first shallow burial predicted burial depth and a first deep burial predicted burial depth; the second predicted burial depth includes a second shallow burial predicted burial depth and a second deep burial predicted burial depth; the first preset neural network and the second preset neural network are different types of neural networks;

[0073] refer to Figure 7 In the depth detection of near-surface defects of tunnel segments, different scale features may be involved due to the different defect depths. For example, shallow defects may correspond to higher-frequency features, while deep defects may correspond to lower-frequency features. To solve this problem, this embodiment uses two neural networks to form a hybrid neural network, enabling them to capture features of different scales, and achieves the fusion of multi-scale information through weighted calculation of the results.

[0074] This embodiment performs multi-scale defect depth calculation on experimental data based on the first preset neural network ResNeXt model and the second preset neural network ConvNeXt V2 model, respectively. Considering the difference between the calculation results of the two models at different scales, an adaptive multi-scale weighted algorithm is proposed to combine the prediction results of the two models:

[0075] First, the acquired defect burial depth signal y is divided into a shallow burial range signal y_l and a deep burial range signal y_h according to the burial depth range. The burial depth range can be divided and selected according to the designed thickness of the tunnel segment test block and the distribution of the designed embedded defects;

[0076] Specifically, in this example, the defect depth signal within the burial depth range of [10, 65] mm is classified as a shallow burial range signal y_l, and the defect depth signal within the burial depth range of [70, 170] mm is classified as a deep burial range signal y_h.

[0077] The defect depth signal y is input into the hybrid neural network model, and the defect depth signal y obtains the corresponding first pre-buried depth y1 through the first preset neural network, and the defect depth signal y obtains the corresponding second pre-buried depth y2 through the second preset neural network. It can be understood that since the defect depth signal y is composed of the shallow buried range signal y_l and the deep buried range signal y_h, when obtaining the first predicted burial depth, the first shallow buried predicted burial depth y_l1 and the first deep buried predicted burial depth y_h1 can be obtained at the same time, and when obtaining the second predicted burial depth, the second shallow buried predicted burial depth y_l2 and the second deep buried predicted burial depth y_h2 can be obtained at the same time.

[0078] Specifically, the first preset neural network obtains the corresponding first predicted burial depth according to the input defect burial depth signal; the first preset neural network is trained by obtaining the first time-frequency image data according to the short-time Fourier transform of the defect burial depth signal;

[0079] The Short-Time Fourier Transform (STFT) is implemented by sliding a fixed-length time window on the signal and performing Fourier transform on the signal fragments within the window. The core idea of ​​this method is to divide the signal into a series of short-time fragments and then perform frequency analysis on these fragments separately. Since each fragment is extracted from the original signal, they can capture the frequency components of the signal near a specific time point.

[0080] After the defect depth signal is transformed by short-time Fourier transform, a series of time-frequency spectrum images will be obtained, such as Figure 8 As shown in the figure, in the time-frequency spectrogram, the horizontal axis usually represents time, the vertical axis represents frequency, and the color or brightness represents the frequency intensity or energy at a specific time point. These frequency graphs show the frequency components of the signal in different time periods, which are called the first time-frequency graphs here;

[0081] According to the test block design of the simulated tunnel segment, the actual collected data defect depth ranges from 10mm to 170mm, and the defect depth moves down 5mm each time. 100 data of each defect signal at each depth are taken for STFT transformation, and 80% of the data are randomly selected for training and 20% of the data are used for verification.

[0082] The first preset neural network is a ResNeXt model;

[0083] Generally, the method of deepening or widening the network is adopted to improve the accuracy of the network model. However, as the number of hyperparameters in the network structure increases (such as the number of channels, filter size, etc.), the difficulty of network design and the computational overhead will also increase. ResNeXt is an improved deep learning architecture that can improve the accuracy without increasing the complexity of parameters, while reducing the number of hyperparameters. It uses the split-transform-merge strategy of Inception, while keeping in mind the concept of VGG repeating blocks with the same structure and Resnet cross-layer branches. The difference between ResNeXt and ResNet-50 is only in the blocks, including BN and ReLU. The ResNeXt block structure is as follows: Fig. 9 shown.

[0084] ResNeXt block has multiple forms of representation, including original form, Concat form, and grouped convolution form. Fig. 9(a) shows the original representation of the ResNeXt block, which uses the split-transform-merge strategy of Inception to construct the corresponding block: reduce the number of channels, run 3X3 convolution, resize the width, and add the results of each branch, where the number of branches is called cardinality, which is 32 in the figure above. Cardinality is regarded as another dimension of the neural network (ResNeXt) in addition to depth and width. Fig. 9 (b) Similar to the original representation, only the merging method is converted to Concat form. Fig. 9 (c) is the simplest form of the ResNeXt block and is also the ResNeXt form used in this embodiment. The model starts with a 1X 1 convolution, then concats the number of channels, and finally applies a 1X 1 convolution to reduce the input 256 channels to 128 channels. Although these three ResNeXt blocks differ in form, they are equivalent.

[0085] In this embodiment, the first preset neural network model adopts the ResNeXt model with a 3*3 convolution kernel. This hyperparameter directly affects the performance of the model, the size of the receptive field, the computational complexity, and the ability to extract features. ResNeXt uses a small convolution kernel, which can better capture local details and texture information and easily extract small changes in the image;

[0086] The first preset neural network is trained using the first time-frequency image data obtained in advance, and the training process is as follows: input the first time-frequency image, normalize the time-frequency image, input the normalized image data into the first preset neural network, and output the corresponding first predicted burial depth.

[0087] Specifically, the second preset neural network obtains the corresponding second predicted burial depth according to the input defect burial depth signal; the second preset neural network is trained by obtaining the second time-frequency image data according to the defect burial depth signal according to S transformation;

[0088] S transform, also known as Stockwell transform, is an extension of the idea of ​​continuous wavelet transform and can also be regarded as a wavelet transform after phase correction. It overcomes the defect of fixed window function width of short-time Fourier transform and uses a movable, scalable and frequency-varying Gaussian window function. Fig.10 .

[0089] The S transform is defined as:

[0090]

[0091] Where: w(τ-t,f) is the Gauss window; τ is the parameter that controls the position of the Gauss window on the time t axis; f is the frequency; x(t) is the audio data collected in the experiment; j is an imaginary unit with a value of e -j2πft is a complex exponential function of frequency f; dt represents the integral of the entire signal x(t) in the time domain;

[0092] in Where σ is The abbreviation of , σ(f) is the scale factor of the Gauss window function;

[0093] According to the test block design of the simulated tunnel segment, the actual collected data defect depth ranges from 10mm to 170mm, and the defect depth moves down 5mm each time. 100 data of each defect signal at each depth are taken for S transformation, and 80% of the data are randomly selected for training and 20% of the data are used for verification.

[0094] ConvNeXt V2 is a new convolutional neural network architecture that integrates the fully convolutional masked autoencoder framework and the global response normalization (GRN) layer in the model architecture, and uses self-supervised learning technology to improve the generalization ability and efficiency of the model.

[0095] In this embodiment, the second preset neural network model is the ConvNeXt V2 model, which uses a 7*7 convolution kernel. This hyperparameter directly affects the performance of the model, the size of the receptive field, the computational complexity, and the ability to extract features. ConvNeXt V2 uses a large convolution kernel, which better captures global features, can cover a larger receptive field, and helps understand the overall structure and contextual information of the image.

[0096] The second preset neural network is a ConvNeXt V2 model;

[0097] The second preset neural network is trained using the second time-frequency image data obtained in advance, and the training process is as follows: inputting the second time-frequency image, normalizing the time-frequency image, inputting the normalized image data into the second preset neural network, and outputting the corresponding second predicted burial depth;

[0098] Before the first preset neural network and the second preset neural network are trained, the method further includes: converting the first time-frequency image data and the second time-frequency image data into a tensor format and performing standardization processing according to formula (1).

[0099] A series of standardized preprocessing steps are performed on the input data processed tunnel segment defect data images, namely the first time-frequency image data and the second time-frequency image data, to ensure that each image received by the model has the same scale and distribution. Such processing helps the model to learn and generalize better, helps the model training to be smoother, and makes the model easier to converge. This includes converting the image data into a tensor format and standardizing it using the following formula (1):

[0100]

[0101] Among them, normalizedpixel is the pixel value after normalization of the image data; pixel is the original pixel value in the image data; mean is the mean vector (0.485, 0.456, 0.406); std is the standard deviation vector (0.229, 0.224, 0.225).

[0102] The mean and std values ​​come from ImageNet, the largest image dataset currently available. This set of mean and std is applicable to most images.

[0103] S3, calculating the average error of the defect burial depth signal, the shallow burial range signal and the deep burial range signal in the first preset neural network and the second preset neural network according to the average value of the difference between the first predicted burial depth and the second predicted burial depth and the actual burial depth;

[0104] The specific steps include:

[0105] S31, respectively calculating the difference between the first predicted burial depth y1 and the second predicted burial depth y2 and the actual burial depth, and then averaging the differences to obtain a first error average value and the second error mean

[0106] S32, respectively calculate the difference between the first shallow buried predicted burial depth y_l1 and the second shallow buried predicted burial depth y_l2 and the actual burial depth, and then average the differences to obtain the first shallow buried error average value and the second shallow buried error average

[0107] S33, respectively calculating the difference between the first deep burial predicted depth y_h1 and the second deep burial predicted depth y_h2 and the actual burial depth, and then averaging the differences to obtain the first deep burial error average value and the average value of the second deep buried error

[0108] This embodiment also provides a specific implementation method for calculating the difference between the predicted burial depth and the actual burial depth based on the model.

[0109] Specifically, taking the ResNeXt model as an example, the data needs to be preprocessed before using the ResNeXt model for burial depth detection. First, for the defect burial depth corresponding to the image, the experiment uses formula (2) to perform maximum-minimum normalization on the predicted burial depth, and maps the burial depth value to between 0 and 1. Formula (2) is as follows:

[0110]

[0111] where d normalized is the normalized burial depth; d min is the minimum burial depth; d max is the maximum burial depth;

[0112] In this experiment, the minimum buried depth of the collected data is d min Set to 10, the maximum burial depth d max Set to 170. The loss function used in the model training process is the mean square error (MSE) loss, which measures the average square difference between the predicted burial depth value and the actual burial depth value, as shown in formula (3):

[0113]

[0114] Where MSE is the mean square error; is the predicted burial depth; i is the actual burial depth value; n is the number of data;

[0115] Each model training lasts for 100 cycles, and an early stopping strategy is used. When the loss on the validation set does not decrease further within 8 consecutive cycles, the training will stop early. This helps to avoid overtraining the model and save the model state with the best validation performance when appropriate. After each training cycle, the performance of the model on the validation set is evaluated by calculating the root mean square error (RMSE), which is then denormalized to the range of the original burial depth value to obtain the actual error of the model in predicting the burial depth, as shown in formula (4):

[0116] RMSE original =RMSE normalized ×(d max -d min )(4)

[0117] Among them, RMSE original Represents the root mean square error after denormalization, that is, the root mean square error on the original data set; RMSE normalized is the root mean square error obtained after normalizing the data;

[0118] Fig.11 This is the ResNeXt model burial depth prediction loss change curve.

[0119] It is understandable that the above method can also be used in the ConvNeXt V2 model to predict the change of the buried depth difference and the change of the buried depth loss.

[0120] S4. Reciprocally normalize the error averages of the defect depth signal, shallow burial range signal and deep burial range signal in the first preset neural network and the second preset neural network to obtain their weight ratios in the two neural networks, and then calculate and output the final burial depth according to the adaptive multi-scale weighted algorithm.

[0121] The specific steps include:

[0122] S41, reciprocally normalizing the error averages of the defect depth signal in the first preset neural network and the second preset neural network to obtain a first depth weight ratio Q1 and a second depth weight ratio Q2;

[0123] Calculate the first error mean and the second error mean The reciprocal of By normalizing these inverses, we can get the weight ratio of the defect depth signal in the results of the two models. The first error weight Second error weight

[0124] S42, reciprocally normalizing the error averages of the shallow burial range signal in the first preset neural network and the second preset neural network to obtain a first shallow burial weight ratio Q_l1 and a second shallow burial weight ratio Q_l2;

[0125] Calculate the average value of the first shallow burial error and the second shallow buried error average The reciprocal of By normalizing these inverses, we can get the weight ratio of shallow buried range signals in the results of the two models. The first shallow buried error weight is Second shallow buried error weight

[0126] S43, reciprocally normalizing the error averages of the deep burial range signal in the first preset neural network and the second preset neural network to obtain a first deep burial weight ratio Q_h1 and a second deep burial weight ratio Q_h2;

[0127] Calculate the average value of the first depth error and the average value of the second deep buried error The reciprocal of And normalizing these inverses can get the weight ratio of the deep buried range signal in the results of the two models, the first deep buried error weight, the second deep buried error weight

[0128] S44, calculating and outputting the final burial depth according to an adaptive multi-scale weighted algorithm; the adaptive multi-scale weighted algorithm is as follows:

[0129]

[0130] Where y is the final burial depth; thr is the preset burial depth threshold, thr = (d max -d min )÷2+d min , thr is a relative number, which is determined according to the experimental design and has different values ​​in different situations. Preferably, in this embodiment, thr is 90, and the unit is mm.

[0131] From the results, it can be seen that the average absolute value of the difference between the calculation results of all collected data of the test block simulating the tunnel segment on the ResNeXt depth prediction model and the actual burial depth is 4.782mm, the average absolute value of the difference between the calculation results of the y_l data on the ResNeXt depth prediction model and the actual burial depth is 3.863mm, and the average absolute value of the difference between the calculation results of the y_h data on the ResNeXt depth prediction model and the actual burial depth is 5.701mm; the average absolute value of the difference between the calculation results of all collected data of the test block on the ConvNeXt depth prediction model and the actual burial depth is 4.613mm, the average absolute value of the difference between the calculation results of the y_l data on the ConvNeXt depth prediction model and the actual burial depth is 5.131mm, and the average absolute value of the difference between the calculation results of the y_h data on the ConvNeXt depth prediction model and the actual burial depth is 4.095mm.

[0132] Table 1 Weight calculation

[0133] data ResNeXt ConvNeXt Weight Q y_l 3.863 5.131 Q_l1=57.05%, Q_l2=42.95% Y_h 5.701 4.095 Q_h1=41.80%, Q_h2=58.20% All data 4.782 4.613 Q1=49.10%,Q2=50.90%

[0134] According to the setting of the hybrid network model, in the subsequent test network, the weight ratios of ResNeXt and ConvNeXt are shown in Table 2. New data are randomly collected on the test block, and the data of known defect depths that are not used for model training are processed and input into the trained model for burial depth calculation, and then the calculation results are output. It is found that most of the errors in the judgment of each burial depth are in the range of 1mm to 4mm, as shown in Table 2.

[0135] Table 2 Calculation error of some burial depths

[0136]

[0137]

[0138] It can be seen from the data in the table that the re-collected depth data performs well on the hybrid model, with an average error of only 1.76 mm. Compared with the prediction error values ​​of each network model individually, the calculation results of the hybrid depth prediction model proposed in this paper are better.

[0139] The tunnel segment near-surface defect detection method based on hybrid deep learning network provided in this embodiment includes the following steps: S1, using the impact echo method to collect the near-surface defect burial depth signal of the tunnel segment test block; S2, inputting the defect burial depth signal into the first preset neural network and the second preset neural network to obtain the first predicted burial depth and the second predicted burial depth; S3, calculating the error average value of the defect burial depth signal, the shallow buried range signal and the deep buried range signal in the first preset neural network and the second preset neural network according to the average value of the difference between the first predicted burial depth and the second predicted burial depth and the actual burial depth; S4, respectively reciprocally normalizing the error average values ​​of the defect burial depth signal, the shallow buried range signal and the deep buried range signal in the first preset neural network and the second preset neural network to obtain their weight ratios in the two neural networks, and then calculating and outputting the final burial depth according to the adaptive multi-scale weighted algorithm. This embodiment mainly focuses on using the impact echo method based on the air-coupled ultrasonic array to generate and collect acoustic wave data, and converting the acoustic wave data into image data through the data processing process, using the hybrid neural network to calculate the burial depth of the defect, realizing the detection of near-surface defects of the tunnel segment, and improving the accuracy and efficiency of tunnel segment detection.

[0140] In addition, the average absolute value of the difference between the calculation result of the data collected in this embodiment on the hybrid depth prediction model and the actual burial depth is 1.76 mm, which can well locate the position of the near-surface defects of the tunnel segment and has guiding significance for actual engineering practice;

[0141] In addition, in this implementation, a series of standardized preprocessing steps were performed on the tunnel segment defect data images before model training to ensure that each image received by the model has the same scale and distribution. Such processing helps the model to learn and generalize better, helps the model training to be smoother, and makes the model easier to converge.

[0142] In addition, this embodiment provides a method for making tunnel segment test blocks, which can simulate the actual situation of tunnel segments under controlled laboratory conditions and allow researchers to accurately control the experimental conditions, making the data more controllable, and the production of laboratory test blocks can follow a standardized process, making the experimental results more comparable.

[0143] This specification is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of this specification. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 Processes Multiple processes and / or boxes Figure 1 A device that has the functions specified in one or more boxes.

[0144] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0145] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0146] The parts of the present invention that are not described in detail are prior art. It is obvious to those skilled in the art that the present invention is not limited to the details of the above-mentioned exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention; therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive, and it is intended that all changes that fall within the meaning and scope of equivalent elements are included in the present invention.

Claims

1. A tunnel segment near-surface defect detection method based on a hybrid deep learning network, characterized in that: The following steps are involved: S1. Use the impact echo method to collect the buried depth signal of the near-surface defects of the tunnel segment test block; S2, inputting the defect burial depth signal into the first preset neural network and the second preset neural network to obtain the first predicted burial depth and the second predicted burial depth; the defect burial depth signal includes a shallow burial range signal and a deep burial range signal; the first predicted burial depth includes a first shallow burial predicted burial depth and a first deep burial predicted burial depth; the second predicted burial depth includes a second shallow burial predicted burial depth and a second deep burial predicted burial depth; the first preset neural network and the second preset neural network are different types of neural networks; S3, calculating the average error of the defect burial depth signal, the shallow burial range signal and the deep burial range signal in the first preset neural network and the second preset neural network according to the average value of the difference between the first predicted burial depth and the second predicted burial depth and the actual burial depth; S4. Reciprocally normalize the error averages of the defect depth signal, shallow burial range signal and deep burial range signal in the first preset neural network and the second preset neural network to obtain their weight ratios in the two neural networks, and then calculate and output the final burial depth according to the adaptive multi-scale weighted algorithm.

2. The method according to claim 1, characterized in that Step S3 specifically includes the following steps: S31, respectively calculating the difference between the first predicted burial depth y1 and the second predicted burial depth y2 and the actual burial depth, and then averaging the differences to obtain a first error average value and the second error mean S32, respectively calculate the difference between the first shallow buried predicted burial depth y_l1 and the second shallow buried predicted burial depth y_l2 and the actual burial depth, and then average the differences to obtain the first shallow buried error average value and the second shallow buried error average S33, respectively calculating the difference between the first deep burial predicted depth y_h1 and the second deep burial predicted depth y_h2 and the actual burial depth, and then averaging the differences to obtain the first deep burial error average value and the average value of the second deep buried error 3. The method according to claim 1, characterized in that Step S4 specifically includes the following steps: S41, reciprocally normalizing the error averages of the defect depth signal in the first preset neural network and the second preset neural network to obtain a first depth weight ratio Q1 and a second depth weight ratio Q2; S42, reciprocally normalizing the error averages of the shallow burial range signal in the first preset neural network and the second preset neural network to obtain a first shallow burial weight ratio Q_l1 and a second shallow burial weight ratio Q_l2; S43, reciprocally normalizing the error averages of the deep burial range signal in the first preset neural network and the second preset neural network to obtain a first deep burial weight ratio Q_h1 and a second deep burial weight ratio Q_h2; S44, calculating and outputting the final burial depth according to an adaptive multi-scale weighted algorithm; the adaptive multi-scale weighted algorithm is as follows: Where y is the final burial depth; thr is the preset burial depth threshold, thr = (d max -d min )÷2+d min , d min is the minimum burial depth; d max is the maximum burial depth.

4. The method according to claim 1, characterized in that: Step S1 specifically includes: performing impact echo detection on the surface of the tunnel segment test block, and then collecting acoustic wave data through an air-coupled probe to obtain a defect burial depth signal.

5. The method according to claim 1, characterized in that Step S1: Defects of the same type and different burial depths are buried in the tunnel segment test block; the defects are filled with hard foam materials.

6. The method according to claim 1, characterized in that In step S2, the first preset neural network obtains the corresponding first predicted burial depth according to the input defect burial depth signal; the first preset neural network is trained by obtaining the first time-frequency image data according to the short-time Fourier transform of the defect burial depth signal.

7. The method according to claim 1, characterized in that In step S2, the first preset neural network is a ResNeXt model.

8. The method according to claim 1, characterized in that In step S2, the second preset neural network obtains the corresponding second predicted burial depth according to the input defect burial depth signal; the second preset neural network is trained by obtaining the second time-frequency image data according to the S transform of the defect burial depth signal.

9. The method according to claim 1, characterized in that: In step S2, the second preset neural network is a ConvNeXt V2 model.

10. The method according to claim 1, characterized in that Before the first preset neural network and the second preset neural network are trained, the method further includes: converting the first time-frequency image data and the second time-frequency image data into a tensor format and performing standardization processing through formula (1), which is as follows: Among them, normalizedpixel is the pixel value after normalization of the image data; pixel is the original pixel value in the image data; mean is the mean vector; std is the standard deviation vector.

Images