A single-stage lightweight subway lining cavity identification method and related equipment
By performing feature extraction and fusion optimization on the YOLOv7 detection model, and using ground penetrating radar deep convolution to generate synthetic data training for adversarial network optimization, the problem of insufficient accuracy in tunnel hole detection is solved, and more accurate and Robust detection results to ensure tunnel safety.
Patent Information
- Application Number
- CN202311102838.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-08-29
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2043-08-29
AI Technical Summary
In the prior art, there is insufficient accuracy when using a single-stage method to detect tunnel void disease data, resulting in inaccurate detection results.
The YOLOv7 detection model is used for feature extraction and fusion optimization, and the synthetic data obtained from the optimization of the adversarial network is trained through real data and ground penetrating radar deep convolution to enhance the model's ability to extract local and global features, reduce computing costs, and improve the model's robustness.
It improves the accuracy and robustness of the detection model, can provide effective and accurate detection results in a short time, discover potential voids in the tunnel lining, prevent accidents caused by voids, and ensure the safety of the tunnel internally.
Smart Images

Figure CN117351321B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of tunnel inspection and maintenance, and in particular to a single-stage lightweight subway lining cavity identification method system, an intelligent terminal and a computer-readable storage medium. Background Art
[0002] Regular inspection and maintenance of subway tunnel linings are critical to ensuring safe and reliable operation of underground subway systems. Although tunnel linings are built with durable materials, they degrade over time due to various factors such as wear, water intrusion and environmental conditions. Early detection of potential problems is crucial to preventing service disruptions, safety hazards or catastrophic events. However, traditional vision and lidar inspection methods can only inspect the surface of the lining. A penetrating inspection method is needed to detect defects inside the lining, such as cracks, voids and delamination.
[0003] The formation of voids in tunnel linings is mainly due to insufficient grouting between lining layers. These voids are usually thin and elongated in both vertical and horizontal dimensions. In addition, some holes in the lined rock are circular in shape and are larger and deeper due to soil erosion. When voids appear in tunnel linings, the difference in dielectric constant between the air inside the void and the surrounding concrete produces strong reflections, which appear as high amplitudes in radar images.
[0004] Ground Penetrating Radar (GPR) is a non-destructive testing method widely used in tunnel lining inspection due to its strong penetration ability, high resolution, high efficiency and non-contact. However, the interpretation of tunnel void disease data, namely GPR data, is challenging due to its reliance on professional knowledge and experience, resulting in differences in the interpretation results, which affects the accuracy and consistency of tunnel lining void detection. In addition, the infrastructure, reinforcement materials and uneven lining materials in the tunnel will generate complex noises, which will interfere with and distort the response of the ground penetrating radar, leading to human bias and misunderstanding. At present, the deep learning framework includes two types of GPR interpretation: a two-stage method including recognition and classification, and a single-stage method. The single-stage method is lightweight and more suitable for analyzing large amounts of data using portable devices, which is convenient for interpreting GPR data. Among the single-stage methods, the You Only Look Once (YOLO detection model, referred to as YOLO) method is more effective in interpreting GPR data.
[0005] However, the single-stage method in the prior art cannot provide the same accuracy as the two-stage method. When detecting tunnel cavity disease data, the use of the single-stage method still has shortcomings, resulting in inaccurate detection results.
[0006] Therefore, the existing technology still needs to be improved and developed. Summary of the invention
[0007] The main purpose of the present invention is to provide a single-stage lightweight subway lining void identification method system, an intelligent terminal and a computer-readable storage medium, aiming to solve the problem that in the prior art, when detecting tunnel void disease data, the use of a single-stage method still has shortcomings, resulting in inaccurate detection results.
[0008] In order to achieve the above object, the first aspect of the present invention provides a single-stage lightweight subway lining cavity identification method, wherein the single-stage lightweight subway lining cavity identification method comprises:
[0009] Acquire tunnel cavity disease data, and perform tunnel cavity preprocessing on the tunnel cavity disease data to obtain preprocessed data, wherein the tunnel disease data at least includes tunnel cavity location information and tunnel cavity size information;
[0010] The preprocessed data is input into a trained optimization detection model to obtain and output a picture with a hole detection result, wherein the optimization detection model is trained by real data and synthetic data optimized by a ground penetrating radar deep convolution generative adversarial network.
[0011] Optionally, the acquiring of tunnel cavity disease data and performing tunnel cavity preprocessing on the tunnel cavity disease data to obtain preprocessed data specifically includes:
[0012] Acquiring the tunnel cavity disease data by radar;
[0013] The tunnel cavity disease data is subjected to tunnel cavity preprocessing to obtain the preprocessed data, wherein the preprocessing includes at least one of direct current offset elimination, gain adjustment, time zero point correction, bandpass filtering and sliding average operation.
[0014] Optionally, the step of generating the optimized detection model includes:
[0015] Based on the YOLOv7 detection model, the optimized detection model is obtained by optimizing the feature extraction and fusion process and optimizing the loss function.
[0016] Optionally, the above feature extraction and fusion process is optimized, specifically including:
[0017] The above feature extraction and fusion processes are optimized through pure convolution modules, self-attention and convolution fusion modules, and optimized efficient aggregation networks.
[0018] Optionally, the above optimization of the loss function specifically includes:
[0019] The above loss function is optimized by weighted intersection-over-union function.
[0020] Optionally, the above-mentioned optimized detection model is trained by using real data and synthetic data optimized by a ground penetrating radar deep convolution generative adversarial network, specifically including:
[0021] Preliminary synthetic data were generated based on the finite-difference time-domain method;
[0022] The above-mentioned preliminary synthetic data is optimized according to the trained ground penetrating radar deep convolutional generative adversarial network to obtain the above-mentioned synthetic data.
[0023] Optionally, the above-mentioned ground penetrating radar deep convolutional generative adversarial network includes a generator network and a discriminator network. The above-mentioned ground penetrating radar deep convolutional generative adversarial network is trained by the generator network and the discriminator network to obtain the above-mentioned trained ground penetrating radar deep convolutional generative adversarial network.
[0024] A second aspect of the present invention provides a single-stage lightweight subway lining cavity identification system, wherein the single-stage lightweight subway lining cavity identification system comprises:
[0025] A data preprocessing module, used for acquiring tunnel cavity disease data, and performing tunnel cavity preprocessing on the tunnel cavity disease data to obtain preprocessed data, wherein the tunnel disease data at least includes tunnel cavity location information and tunnel cavity size information;
[0026] The data processing module is used to input the preprocessed data into the trained optimization detection model to obtain and output a picture with a hole detection result, wherein the optimization detection model is trained by real data and synthetic data obtained by optimizing the ground penetrating radar deep convolution generative adversarial network.
[0027] A third aspect of the present invention provides an intelligent terminal, which includes a memory, a processor, and a single-stage lightweight subway lining void identification program stored in the memory and executable on the processor. When the single-stage lightweight subway lining void identification program is executed by the processor, any step of the single-stage lightweight subway lining void identification method is implemented.
[0028] A fourth aspect of the present invention provides a computer-readable storage medium, on which a single-stage lightweight subway lining void identification program is stored. When the single-stage lightweight subway lining void identification program is executed by a processor, any step of the single-stage lightweight subway lining void identification method is implemented.
[0029] As can be seen from the above, in the scheme of the present invention, tunnel cavity disease data is obtained, and tunnel cavity preprocessing is performed on the tunnel cavity disease data to obtain preprocessed data, wherein the tunnel disease data at least includes tunnel cavity location information and tunnel cavity size information; the preprocessed data is input into the trained optimization detection model to obtain and output a picture with cavity detection results, wherein the optimization detection model is trained by real data and synthetic data obtained by optimizing the adversarial network generated by deep convolution of the ground penetrating radar.
[0030] Compared with the prior art, in view of the fact that the single-stage method is still insufficient when detecting tunnel cavity disease data, resulting in inaccurate detection results, the single-stage lightweight subway lining cavity recognition method proposed in the present application processes the tunnel cavity disease data by using the optimized detection model obtained by optimizing the feature extraction and fusion process and loss function of the YOLOv7 (one of the YOLO detection models, referred to as YOLOv7) detection model, thereby enhancing the process of extracting local features, global features and fusion features of the optimized detection model, and at the same time reducing the computational cost without affecting the performance of the model, thereby improving the accuracy of the model, and at the same time optimizing the loss function to improve the robustness of the model; and in the present application, the optimized detection model of the model The training data used in training is synthetic data obtained by optimizing the real data and the deep convolution generative adversarial network of the ground penetrating radar, which makes the generated synthetic data more formal. In addition, since the combination of synthetic data and real data is used to train the optimized detection model, the number of training data sets is greatly increased, thereby improving the training effect of the optimized detection model, and a more accurate trained optimized detection model can be obtained. The obtained optimized detection model can process the tunnel cavity disease data obtained by the ground penetrating radar in a short time to obtain effective and accurate detection results. The detection results can be used to find out where the potential cavities in the tunnel lining appear, so that corresponding processing can be carried out to prevent accidents caused by cavities in the tunnel, thereby making the tunnel safer. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments or the description of the prior art 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 labor.
[0032] Figure 1 It is a flow chart of a single-stage lightweight subway lining cavity identification method provided by an embodiment of the present invention;
[0033] Figure 2 Schematic diagram of the network architecture of YOLOv7 provided by an embodiment of the present invention;
[0034] Figure 3 Schematic diagram of the specific structure of the pure convolution module provided by an embodiment of the present invention;
[0035] Figure 4 is a schematic diagram of the structure of an optimized detection model provided by an embodiment of the present invention;
[0036] Figure 5 It is a schematic diagram of the structure of the fusion module of self-attention and convolution provided in an embodiment of the present invention;
[0037] Figure 6 It is a schematic diagram of the structure of the parallel convolution and optimized efficient aggregation network provided by an embodiment of the present invention;
[0038] Figure 7 is a schematic diagram of simulating a cavity using three different shapes provided by an embodiment of the present invention;
[0039] Figure 8 is a schematic diagram of preliminary synthetic data provided by an embodiment of the present invention;
[0040] Fig. 9 It is a schematic diagram of the training process of the ground penetrating radar deep convolution generative adversarial network provided by an embodiment of the present invention;
[0041] Fig.10 is a schematic diagram of a data collection process provided by an embodiment of the present invention;
[0042] Fig.11 is a schematic diagram of a tunnel structure provided by an embodiment of the present invention;
[0043] Fig.12 It is a schematic diagram of component modules of a single-stage lightweight subway lining cavity identification system provided by an embodiment of the present invention;
[0044] Fig.13 It is a block diagram of the internal structure principle of an intelligent terminal provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0045] In the following description, specific details such as specific system structures and technologies are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present invention. However, it should be clear to those skilled in the art that the present invention may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present invention.
[0046] It should be understood that when used in this specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.
[0047] It should also be understood that the terms used in the present specification are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the present specification and the appended claims, the singular forms "a", "an" and "the" are intended to include plural forms unless the context clearly indicates otherwise.
[0048] It should be further understood that the term "and / or" used in the present description and the appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.
[0049] As used in this specification and the appended claims, the term "if" may be interpreted as "when" or "upon" or "in response to determining" or "in response to being classified into," depending on the context. Similarly, the phrase "if it is determined" or "if classified into [described condition or event]" may be interpreted as meaning "upon determination" or "in response to determining" or "upon classification into [described condition or event]" or "in response to being classified into [described condition or event]," depending on the context.
[0050] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0051] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.
[0052] Regular inspection and maintenance of subway tunnel linings are critical to ensuring safe and reliable operation of underground subway systems. Although tunnel linings are built with durable materials, they degrade over time due to various factors such as wear, water intrusion and environmental conditions. Early detection of potential problems is crucial to preventing service disruptions, safety hazards or catastrophic events. However, traditional vision and lidar inspection methods can only inspect the surface of the lining. A penetrating inspection method is needed to detect defects inside the lining, such as cracks, voids and delamination.
[0053] Ground Penetrating Radar (GPR) is a non-destructive testing method widely used in tunnel lining inspection due to its strong penetration, high resolution, high efficiency and non-contact. GPR exploration relies on the propagation and reflection of electromagnetic waves in tunnel linings. The GPR transmitting antenna transmits electromagnetic waves with a center frequency through the antenna to the target stratum, propagating in the matrix material. When encountering a surface with different electromagnetic properties, the wave is reflected or refracted back and recorded by the receiver. The reflection amplitude is determined by the contrast of dielectric constants, while the conductivity attenuates the GPR signal as it penetrates downward, and the superimposed signals construct a radar map.
[0054] The formation of voids in tunnel linings is mainly due to insufficient grouting between lining layers. These voids are usually thin and elongated in both vertical and horizontal dimensions. In addition, some holes in the lined rock are circular in shape and are larger and deeper due to soil erosion. When voids appear in tunnel linings, the difference in dielectric constant between the air inside the void and the surrounding concrete produces strong reflections, which appear as high amplitudes in radar images.
[0055] The interpretation of GPR data is challenging due to the reliance on expertise and experience, as well as the clutter generated by the distortion and interference of voids in the tunnel lining by the steel bars above. This leads to differences in the interpretation results obtained by different methods, which affects the accuracy and consistency of tunnel lining void detection. In addition, the infrastructure, reinforcement materials and uneven lining materials in the tunnel generate complex noise that interferes with and distorts the response of the ground penetrating radar, leading to human bias and misunderstanding. As the length of subway tunnels increases, the manual interpretation of GPR data becomes more and more time-consuming, and a fast and stable automatic GPR interpretation method needs to be developed.
[0056] Currently, deep learning-based GPR interpretation has recently emerged as an effective solution to this problem. Deep learning frameworks have been divided into two types of GPR interpretation: a two-stage approach that includes recognition and classification, and a single-stage approach. There are currently several methods that use a two-stage approach to facilitate automatic tunnel lining GPR interpretation, including tunnel lining defect segmentation using convolutional neural networks, tunnel lining element identification using mask area-based neural networks, etc. However, these methods are computationally demanding, which poses a challenge to operating servers. In contrast, single-stage methods are lightweight and more suitable for analyzing large amounts of data using portable devices. Among popular single-stage methods such as RetinaNet, DetecNet, and SqueezeNet, the You Only Look Once method has proven to be more effective in interpreting GPR data.
[0057] However, currently the single-stage method does not provide comparable accuracy to the two-stage method, so there is a need to improve the accuracy of the single-stage method while maintaining its computational efficiency. Moreover, compared with optical images, target detection in GPR radar images is more challenging. This is because the superimposed and diffracted subsurface signals make it difficult to identify target objects, resulting in unclear boundaries and features. In addition, the low contrast between different materials in the radar image, coupled with the subtle differences in the dielectric properties of the subsurface materials, further complicates the distinction between objects with similar properties, making the radar image noisy and blurry. Therefore, it is necessary to improve the feature extraction capability of the single-stage method to improve its accuracy.
[0058] In order to solve at least one of the above-mentioned problems, the scheme of the present invention obtains tunnel cavity disease data, performs tunnel cavity preprocessing on the tunnel cavity disease data to obtain preprocessed data, wherein the tunnel disease data at least includes tunnel cavity location information and tunnel cavity size information; the preprocessed data is input into a trained optimization detection model to obtain and output a picture with cavity detection results, wherein the optimization detection model is trained by real data and synthetic data obtained by optimizing a ground penetrating radar deep convolution generative adversarial network.
[0059] Compared with the prior art, in view of the fact that the single-stage method is still insufficient when detecting tunnel cavity disease data, resulting in inaccurate detection results, the single-stage lightweight subway lining cavity recognition method proposed in the present application processes the tunnel cavity disease data by using the optimized detection model obtained by optimizing the feature extraction and fusion process and loss function of the YOLOv7 (one of the YOLO detection models, referred to as YOLOv7) detection model, thereby enhancing the process of extracting local features, global features and fusion features of the optimized detection model, and at the same time reducing the computational cost without affecting the performance of the model, thereby improving the accuracy of the model, and at the same time optimizing the loss function to improve the robustness of the model; and in the present application, the optimized detection model of the model The training data used in training is synthetic data obtained by optimizing the real data and the deep convolution generative adversarial network of the ground penetrating radar, which makes the generated synthetic data more formal. In addition, since the combination of synthetic data and real data is used to train the optimized detection model, the number of training data sets is greatly increased, thereby improving the training effect of the optimized detection model, and a more accurate trained optimized detection model can be obtained. The obtained optimized detection model can process the tunnel cavity disease data obtained by the ground penetrating radar in a short time to obtain effective and accurate detection results. The detection results can be used to find out where the potential cavities in the tunnel lining appear, so that corresponding processing can be carried out to prevent accidents caused by cavities in the tunnel, thereby making the tunnel safer.
[0060] Exemplary Methods
[0061] like Figure 1 As shown, an embodiment of the present invention provides a single-stage lightweight subway lining cavity identification method. Specifically, the single-stage lightweight subway lining cavity identification method comprises the following steps:
[0062] Step S100, acquiring tunnel cavity disease data, and performing tunnel cavity preprocessing on the tunnel cavity disease data to obtain preprocessed data, wherein the tunnel disease data at least includes tunnel cavity location information and tunnel cavity size information.
[0063] Specifically, the tunnel cavity disease data is a radar map obtained by a ground penetrating radar, where the vertical axis of the radar map represents the detection depth. By comparing images at different depths, the location and depth of the underground cavity can be determined. The shape and size of the cavity can be estimated by observing the signal intensity changes in the radar map. Therefore, the tunnel disease data contains tunnel cavity location information and tunnel cavity size information, as well as tunnel cavity shape information. When using a model to process data, the data obtained by the model must first be pre-tunnel cavity processed accordingly.
[0064] Furthermore, the tunnel cavity disease data is obtained, and tunnel cavity preprocessing is performed on the tunnel cavity disease data to obtain preprocessed data, specifically including:
[0065] Acquiring the tunnel cavity disease data by radar;
[0066] The tunnel cavity disease data is subjected to tunnel cavity preprocessing to obtain the preprocessed data, wherein the preprocessing includes at least one of direct current offset elimination, gain adjustment, time zero point correction, bandpass filtering and sliding average operation.
[0067] Specifically, the tunnel cavity disease data is obtained through ground penetrating radar. The ground penetrating radar transmitting antenna transmits electromagnetic waves with a center frequency to the target subway tunnel through the antenna, which propagates in the matrix material. When encountering a surface with different electromagnetic properties, the wave is reflected or refracted back and recorded by the receiver. The reflection amplitude is determined by the comparison of the dielectric constant, and the conductivity will attenuate the GPR signal when penetrating downward. The superimposed signals construct a radar map, that is, the corresponding tunnel cavity disease data is obtained. In one embodiment of the present application, the frequency of the ground penetrating radar transmitting antenna is preferably selected from 500MHz and 800MHz.
[0068] After obtaining the above-mentioned tunnel void disease data, the above-mentioned tunnel void disease data is subjected to tunnel void preprocessing to obtain the above-mentioned preprocessed data. The above-mentioned tunnel void preprocessing includes DC offset elimination, gain adjustment, time zero point correction, bandpass filtering and sliding average operation to obtain preprocessed data for tunnel inspection.
[0069] Among them, DC drift elimination specifically includes: DC drift refers to the DC offset in the signal, that is, the offset of the signal at zero frequency; DC drift elimination is a method of removing the DC offset in the signal, so that the value of the signal at zero frequency is zero, thereby making the signal more stable and accurate.
[0070] Gain adjustment includes: Gain adjustment is to adjust the amplitude of the signal to adapt it to specific needs or ranges; through gain adjustment, the amplitude of the signal can be amplified or reduced so that the signal can be better processed in subsequent processing steps.
[0071] Time zero point correction includes: Time zero point correction is to eliminate time delay or time offset in the signal; by correcting the time zero point, the starting point of the signal can be aligned or corrected to the required time point to ensure the accuracy and consistency of the signal in time.
[0072] Bandpass filtering includes: Bandpass filtering is a signal processing technique that selects signals within a specific frequency range (bandpass) and filters out signals in other frequency ranges; bandpass filtering can be used to remove noise or unwanted frequency components and retain the frequency components of interest.
[0073] Sliding average includes: Sliding average is a method of smoothing signals, which reduces noise and irregular fluctuations in the signal by averaging the signal; Sliding average usually uses a sliding window, calculates the average value of the signal in each window, and uses the average value as the smoothed signal value.
[0074] During preprocessing, the user pre-selects one or more operations to perform tunnel cavity preprocessing according to the specific application scenario, data characteristics and required data quality. After tunnel cavity disease data is preprocessed, the corresponding preprocessed data is obtained. After DC offset elimination, gain adjustment, time zero point correction, bandpass filtering and sliding average operations, the obtained preprocessed data can be optimized for detection model processing. More accurate detection results can be obtained.
[0075] Step S200, input the above-mentioned preprocessed data into the trained optimization detection model to obtain and output a picture with a hole detection result, wherein the above-mentioned optimization detection model is trained by real data and synthetic data obtained by optimizing the ground penetrating radar deep convolution generative adversarial network.
[0076] Specifically, after obtaining the above-mentioned preprocessed data, the preprocessed data is input into the trained optimization detection model, and then a picture with a hole detection result can be output accordingly, wherein the detection result is represented accordingly by an identification box and a confidence rate in the picture with the hole detection result. Through the picture with the hole detection result, it can be known which part of the tunnel has a hole, and it can be repaired quickly accordingly, thereby improving the safety of tunnel use.
[0077] Furthermore, the above-mentioned optimized detection model specifically includes:
[0078] Based on the YOLOv7 detection model, the above-mentioned optimized detection model is obtained by optimizing the feature extraction and fusion process and optimizing the loss function.
[0079] Specifically, the optimized detection model of the embodiment of the present application is improved on the basis of the YOLOv7 (one of the YOLO detection models, referred to as YOLOv7) detection model. Among them, compared with the previous versions of the YOLO series of detection models, YOLOv7 adopts an anchor-based method and uses Silu (obtained by a weighted linear combination of sigmoid activation functions) as the activation function in the convolutional layer module. In addition, it also introduces effective ELAN (Effective Long-range Aggregation Network, referred to as ELAN) modules and REP (Re-parameterization, referred to as REP) modules in the network architecture. The ELAN module achieves effective learning and convergence in deeper networks by controlling the shortest and longest gradient paths. The network architecture of YOLOv7 includes four parts, namely input, backbone network, intermediate layer and output layer, as shown in the following figure. Figure 2 shown.
[0080] like Figure 2As shown in the figure, in the network architecture of YOLOv7, the backbone network is responsible for feature extraction, including multiple CNN (Convolutional Neural Networks, CNN for short) basic modules CBS (Compact Bilinear Pooling, CBS for short), ELAN modules and Max Pooling (MP for short), where the ratio of the number of channels of MP1 and MP2 is different, so MP1 and MP2 are used to distinguish. The CBS module uses conventional convolution and batch generation features, which is different from the convolution layer in the traditional CNN network. The ELAN module enhances the stability and generalization ability of the quantity normalization and SiLU activation function model, while the MP module is used for downsampling operations. The middle layer, i.e. the neck, is the feature fusion part and consists of multiple modules, including spatial pyramid pooling and SPPCSPC (patial pyramid pooling based on convolutional and spatial pooling convolutional, spatial pyramid pooling based on convolution and spatial pooling convolution, referred to as SPPCSPC) modules, CBS modules, MP2 modules and UP (upsample, referred to as UP) modules. Its main purpose is to fuse the features of different scales from the backbone network to better detect objects of different sizes. The SPPCSPC module can extract features of different receptive fields, while the CBS module, MP2 module and UP module are used for feature fusion and compression to reduce parameters and computational complexity. Through the feature fusion of the middle layer, YOLOv7 can detect objects of different sizes more accurately, thereby improving detection accuracy and efficiency. The output layer, i.e. the head, uses three detection heads to output the predicted category probability, confidence score and bounding box coordinates of the target. Each detection head corresponds to a different feature map scale and is used to detect objects of different sizes. Each feature map cell predicts the probability and location of an object and outputs a confidence score for the predicted bounding box, indicating whether there is an object in the bounding box. During training, each feature map cell is assigned to detect one or more objects, and the feature map cell responsible for a specific object predicts the category and location of the object. In YOLOv7, the matching of positive and negative samples is based on calculating the intersection over union (IOU) between the target box and the anchor box. Specifically, the model calculates the IOU of each target box with all anchor boxes and assigns it to the anchor box with the highest IOU. If the IOU value of an anchor box with multiple target boxes is less than a certain threshold, it is considered a negative sample. On the contrary, if the IOU value of an anchor box with the target box is greater than a certain threshold, it is considered a positive sample.Finally, if the IOU value between an anchor box and the target box is between two thresholds, it is considered as an ignored sample and is not included in the calculation of the loss function. This is how the YOLOv7 detection model is trained.
[0081] After optimizing the above YOLOv7 detection model, the above optimized detection model can be obtained.
[0082] Furthermore, the above-mentioned optimization of the feature extraction and fusion process specifically includes:
[0083] The above feature extraction and fusion processes are optimized through pure convolution modules, self-attention and convolution fusion modules, and optimized efficient aggregation networks.
[0084] Among them, the specific structure of the pure convolution module (CNXB) is as follows Figure 3 As shown in the figure, the pure convolution module contains a pure convolution block (ConvNeXt Block). The original ELAN module is improved by the ConvNeXt Block. The four 3×3 convolutions in the ELAN module are replaced by the ConvNeXt Block module, and the ConvNeXt Block provides convenience, so that a CNXB module and a CBS module are used to replace the first ELAN module in the backbone network.
[0085] ConvNeXt Block first extracts and normalizes features through a 7×7 deep convolution layer and a normalization layer; then, the channel size is quadrupled and nonlinearity is introduced through 1×1 convolution and GELU (Gaussian Error Linear Unit, an activation function, referred to as GELU) activation function; subsequently, the channel size is restored to its original size through 1×1 convolution, layer scaling, and Drop Path (regularization) layer, and finally the input and output are added together; among them, in ConvNeXtBlock, there is only one normalization layer located after the deep convolution, which uses layer normalization (Layer Normalization, referred to as LN) as normalization. In addition, ConvNeXt Block reduces the requirements for activation and regularization functions because each module only uses one activation function and one regularization function.
[0086] The location where CNXB is added is as follows Figure 4As shown in the figure, in the backbone network, after passing through 4 CBS modules, a feature map of size 160×160×128 is obtained. In order to capture more feature information and provide more useful information for subsequent tasks, the CXNB module is placed at the end of the CBS module sequence as a special block, and an additional 3x3 convolutional layer CBS is added. This can better process the input features and provide higher quality feature representation at the output layer. In addition, the higher computational efficiency of the CXNB module effectively accelerates the training and inference of the model, making it more feasible in practical applications.
[0087] The fusion module of self-attention and convolution (Attention and Convolution Mix, the fusion of self-attention and convolution, referred to as ACmix), in the embodiment of the present application, Acmix combines the self-attention and convolution paradigms, has the advantages of self-attention and convolution, and has a smaller computational overhead than pure convolution or self-attention. The structure of the ACmix module is as follows Figure 5 As shown in Figure 1, the ACmix structure is divided into two stages: Stage I and Stage II. In Stage I, the input features are convolutionally projected and reshaped into N blocks to obtain an intermediate feature set containing feature maps. In Stage II, the structure is divided into two paths: self-attention and convolution. In the self-attention path, the intermediate feature set is clustered into N groups, each containing 3 feature maps. These 3 feature maps are used as queries, keys, and values, respectively, and are processed according to the traditional multi-head self-attention model. In the convolution path, a lightweight fully connected layer is used to generate N groups, each with K 2 feature maps. For the generated feature maps, they are processed by translation and aggregation, processing the input features in a convolutional manner and collecting information from the local receptive field. Finally, the outputs of the self-attention and convolution paths are summed.
[0088] In ACmix, the standard convolution process is:
[0089]
[0090]
[0091]
[0092] Among them, K p,q represents the kernel weight of the kernel position (p, q); g ij and f ij are the output and input feature tensors corresponding to pixel (i, j), respectively; k is the kernel size; is the output feature of the kernel position (p, q); is the output feature of the kernel position (p, q) after the conversion operator; Represents the matrix To translate, move horizontally units, vertical movement p and q are the original row and column indices, It means k is divided by 2 and rounded up, where k is a constant and square brackets are used to indicate rounding down.
[0093] The multi-head attention mechanism process is:
[0094]
[0095]
[0096] in, are the projection matrices for query, key, and value, respectively; Represents a local pixel area with a spatial range of kc centered at (i, j); The superscript (l) represents the number of layers or levels, which means the query vector at position (i, j) at layer l. represents the key vector of position (a, b) at layer l, Represents the value vector of position (a, b) at layer l. yes The corresponding attention weights of the internal features.
[0097] The convolution part and the self-attention part are fused, that is, the outputs of the self-attention and convolution paths are summed and fused using different weights:
[0098] F out =αF att +βF conv , (6);
[0099] F out 、F att 、F conv are the output features of ACmix, self-attention and convolution respectively; α and β are learnable parameters.
[0100] In YOLOv7, the ELAN module and SPPCSPC module located in the backbone and neck of the network enable the network to focus on important areas that contribute to object detection, thereby improving detection accuracy. In order to enhance the ability of the optimized detection model to represent the characteristics of the voids affected by steel bars, three ACmix modules are introduced on the basis of YOLOv7, such as Figure 4As shown. These ACmix modules are placed after the ELAN and SPPCSPC modules, allowing them to operate on the feature maps generated by these modules. Specifically, in the backbone network, an ACmix module is placed after the second and third ELAN modules based on YOLOv7, and an ACmix module is also placed after the SPPCSPC module in the neck. This setting helps to extract local and global features more effectively, thereby enhancing the network's ability to identify and locate objects. In addition, since the ELAN and SPPCSPC modules have focused the network's attention on the key areas of object detection, the ACmix module can accurately extract valuable feature information, thereby improving the accuracy of object detection.
[0101] The optimized efficient aggregation network (PELAN is used in the embodiment of the present application to represent the optimized efficient aggregation network) uses a combination of parallel convolution (PConv) and 1x1CBS modules to replace the CBS module in the original efficient aggregation network on the basis of the efficient aggregation network, thereby creating a new PELAN module. Parallel convolution reduces redundant calculations and memory accesses at the same time, effectively extracting spatial features. The specific structures of PELAN and PConv are as follows: Figure 6 As shown in Figure 1, the design of PConv takes advantage of the redundancy in the feature map and performs regular convolution on only some input channels without affecting the remaining channels. The floating point operations (FLOPs) of PConv are calculated using the following formula:
[0102]
[0103] PConv FLOPs is the FLOPs of PConv; h, w, c p are the height, width and number of channels of the data respectively; k is the number of rows or columns of the 3D tensor in the filter. From the floating point operation amount, we can know that when When , PConv’s FLOPs is only 1 / 16 of that of traditional convolution, which can better utilize computing resources.
[0104] Specific as Figure 4 As shown in the figure, based on YOLOv7, except for the first ELAN module replaced by CXNB and CBS modules, all other ELAN modules are replaced by PELAN; PELAN can flexibly control the convolution operation range through masks, enabling it to handle inputs of various sizes while reducing computational costs. This leads to improved model generalization and reduced computing resource consumption.
[0105] Understandably, in the YOLOv7 framework, the backbone network plays a vital role in feature extraction, which is essential for accurate object recognition in deep learning. In the backbone network, the backbone of YOLOv7 is improved by adding CXNB modules and CBS modules (with a convolution kernel of 3 and a step size of 1) after the fourth CBS module, and adding ACmix modules at the three connection points between the backbone and the neck, thereby improving the effect of feature extraction. In addition, to further optimize the model, the CBS module (with a convolution kernel of 3 and a step size of 1) is replaced with a PELAN module in the ELAN module of the backbone and the neck, thereby reducing the computational complexity and number of parameters of the model. These improvements make the meaningful content and location in the input image samples of the optimized detection model more concentrated, and feature information can be extracted more effectively even in the presence of steel bar interference, thereby improving detection accuracy.
[0106] Furthermore, the above optimization of the loss function specifically includes:
[0107] The above loss function is optimized by weighted intersection-over-union function.
[0108] Specifically, due to the class imbalance and complex background caused by underground heterogeneity, the GPR interpretation based on deep learning will have unstable convergence. Therefore, in the embodiment of the present application, the original loss function weighted intersection in the YOLOv7 detection model is replaced with an optimized loss function, namely the weighted intersection-over-union function Wise-IoU (Weighted Intersection over Union, referred to as WIoU) to solve these problems.
[0109] WIoU uses a dynamic non-monotonic focusing mechanism to balance the bounding box regression of samples with different qualities, thereby improving the overall model performance. In the YOLO series of models, cross entropy is usually used as a loss function, but it may be harmful to model accuracy when there are unbalanced positive and negative samples. The WIoU in this application considers two key indicators: the intersection of the target box and the true box and the IoU between the predicted box and the true box. WIoU adjusts the IoU between the target box and the true box to solve the problem of sample imbalance, while improving the IoU between the predicted box and the true box to enhance model performance. In an embodiment of the present application, the use of the WIoU loss function can better solve the problem of imbalance between positive and negative samples in tunnel lining GPR images, improve model performance, and produce more accurate detection results.
[0110] Specifically, the WIoU loss is defined as follows:
[0111]
[0112]
[0113]
[0114]
[0115] Among them, Wg and Hg represent the width and height of the minimum bounding box respectively. is the monotonic focusing factor, β is the outlier of the dynamic non-monotonic FM anchor frame, γ is the non-monotonic focusing factor, α and δ are hyperparameters, x gt and gt represents the horizontal and vertical coordinates of the GT box, x and y represent the coordinates of the center point of the prediction box, They represent the loss of three versions of WIoU respectively. Represents the weight of IoU, used for weighted calculation r is a scaling factor calculated from a set of constant parameters (β, δ, α). In some loss functions or evaluation metrics, α and δ are used as hyperparameters to adjust the model training or optimization process. The specific values of these parameters are selected based on experience or experiments to obtain better model performance.
[0116] Furthermore, before the optimized detection model is used to detect the tunnel cavity disease data, the optimized detection model must be trained. The data set used to train the optimized detection model consists of actual real data and synthetic radar map data, i.e., synthetic data. In one embodiment of the present application, before the optimized detection model is trained, the data set is divided into parts of uniform size, with a predetermined pixel size of 800×320; finally, the data set consists of 400 parts, which are divided into two sets, namely, a training set and a validation set, wherein the ratio of the training set to the validation set is 8:2. Moreover, when generating the training set, traditional data enhancement techniques are used to process the real data and the synthetic data by cropping, translation, brightness adjustment, rotation, and mirroring, further enhancing the quality and quantity of the data in the data set, and finally obtaining a data set containing 2,000 cavity disease data, and in addition, blurring and scaling operations are used to process the real data and the synthetic data, thereby simulating radar maps generated by different antennas. These data enhancement techniques help to diversify the data set and enable the model to be better generalized to unseen data. The fragments containing various cavity geometric shapes in the dataset are annotated using lambelimg (a graphic image annotation tool) so that the model can learn to recognize different cavity shapes and sizes. By using the diverse and annotated data in the dataset, the accuracy of the optimized detection model can be improved, resulting in a better optimized detection model.
[0117] Furthermore, the above-mentioned optimized detection model is trained by using real data and synthetic data optimized by the deep convolutional generative adversarial network of the ground penetrating radar, specifically including:
[0118] Preliminary synthetic data were generated based on the finite-difference time-domain method;
[0119] The above-mentioned preliminary synthetic data is optimized according to the trained ground penetrating radar deep convolutional generative adversarial network to obtain the above-mentioned synthetic data.
[0120] Since ground truth is not available, numerical simulation is a popular method to study the response of ground penetrating radar in different defect situations. Finite-Difference Time-Domain (FDTD) method is often used to numerically simulate GPR propagation. The FDTD method discretizes the electromagnetic field on a spatial grid and uses a time-stepping technique to simulate the propagation of electromagnetic waves in the medium. In each time step, Maxwell's equations are discretized to calculate the changes in the electric and magnetic fields in space. Therefore, the FDTD method is computationally efficient for modeling problems with multiple sources or a wide frequency range. GPRMAX is an open source software that uses the FDTD method to simulate the propagation of electromagnetic waves in a user-defined model. Therefore, preliminary synthetic data can be generated by the FDTD method as well as GPRMAX.
[0121] In one embodiment of the present application, according to the subway tunnel lining design standard, the lining is composed of two layers of steel mesh, and the thickness of the secondary lining is 0.3 to 0.4 m. Therefore, in the FDTD model, the total height is 1.2 m, the air space height is 0.2 m, the secondary lining thickness is 0.4 m, the primary lining thickness is 0.2 m, and the remaining 0.4 m is the surrounding rock layer. In order to further enhance the generalization ability of the model, FDTD is used to model the tunnel lining voids, which are three scenarios: concrete without reinforcement, one layer of reinforcement, and two layers of reinforcement. The embedded voids vary in size, shape, and position, such as Figure 7 Considering the requirements of ground penetrating radar tunnel detection on spatial resolution and penetration depth, the designed FDTD model is shown in Table 1.
[0122]
[0123] Table 1. FDTD model parameters
[0124] like Figure 7 As shown, Figure 7 In,∈ r is the relative permittivity; σ is the conductivity (Siemens / meter); μ r is the relative magnetic permeability; σ ris the magnetic loss (ohm / meter); three different shapes are used to simulate voids, namely thin elliptical, circular and rectangular. The elliptical voids are located at a shallower depth and represent the boundaries between layers, while the circular and rectangular voids are buried deeper and represent voids in the rock behind the lining. The material properties and structure of the lining model are consistent with the actual situation.
[0125] By using the FDTD method, preliminary synthetic data, i.e., preliminary synthetic radar chart, can be obtained. In one embodiment of the present application, Figure 8 The figure shows the preliminary synthetic radar map. In the preliminary synthetic radar map, the circular cavity produces hyperbolic reflection, while the rectangular cavity produces cross reflection; the reverberation interference of the reinforcement layer and the lining layer is properly modeled, which leads to the distortion of the cavity response. However, in practice, the real cavity presents irregular shapes, and the regular shapes used in the numerical simulation of GPR propagation using the FDTD method are too idealized, so they need to be optimized before they are used as training samples.
[0126] After obtaining the preliminary synthetic data, the preliminary synthetic data is optimized according to the trained ground penetrating radar deep convolutional generative adversarial network to obtain the synthetic data.
[0127] Furthermore, the above-mentioned ground penetrating radar deep convolutional generative adversarial network includes a generator network and a discriminator network. The above-mentioned ground penetrating radar deep convolutional generative adversarial network is trained through the generator network and the discriminator network to obtain the above-mentioned trained ground penetrating radar deep convolutional generative adversarial network.
[0128] Specifically, the ground penetrating radar-DeepConvolutional Generative Adversarial Networks (GPR-DCGAN) is essentially a deep convolutional generative adversarial network (DCGAN), which is used to optimize the generated preliminary synthetic data in the embodiments of the present application. Among them, the GPR-DCGAN architecture includes two deep convolutional neural networks: a generator and a discriminator. The generator in DCGAN consists of five transposed convolutional layers, which maps low-dimensional noise vectors to high-dimensional image space. On the other hand, the discriminator consists of five conventional convolutional layers, which produces a binary classification result, indicating whether the radar map is real or fake. The confrontation between the generator and the discriminator network enables the generator to gradually generate more realistic radar maps.
[0129] The above-mentioned generated preliminary synthetic data has unreal content, so it is necessary to eliminate the mismatch between the real and synthetic radar images. The untrained GPR deep convolutional generative adversarial network cannot process the preliminary synthetic data well, so the real radar image and the preliminary synthetic data generated by the FDTD method are used to form the GPR deep convolutional generative adversarial network training data, and the GPR deep convolutional generative adversarial network is trained by the GPR deep convolutional generative adversarial network training data. During the training process, initially, the real radar image, that is, the real data is designated as "true", and the random noise, that is, the preliminary synthetic data is designated as "false" by the discriminator; the discriminator then extracts and learns the difference between the real and preliminary synthetic data, including error information, and the above error information includes blur caused by dispersion and noise caused by scattering; then the error information is propagated back to the generator through back propagation, and the generator incorporates these errors into the synthetic radar image to enhance its authenticity. At the same time, the generator and the discriminator are trained and updated together, and the generator tries to improve its output to meet the requirements of the discriminator by introducing noise, while the discriminator enhances its ability to correctly distinguish between real data and generated data. Subsequently, the real and modified synthetic radar images are fed back into the discriminator for the next round of identification. This alternating training process continues until the discriminator designates the modified synthetic radar image as "true", indicating that the synthetic radar image generated by the generator is convincing enough that the discriminator cannot distinguish it from the real radar image. The specific training process of the deep convolutional generative adversarial network for ground penetrating radar is as follows: Fig. 9 shown.
[0130] In one embodiment of the present application, in GPR-DCGAN, the learning rate is fine-tuned to balance the convergence speed and stability, and is finally set to 0.0002. The selected optimizer is "Adam (Adaptive Moment Estimation, Adam optimizer, referred to as Adam)", and the model is trained from scratch without loading any pre-trained weights. The input and output image sizes are set to 800×320 pixels, and the training is stopped when the discriminator loss reaches a preset threshold. In one embodiment, the preset threshold is set to stop training when it approaches 0. At this time, the weights of the generator are used to optimize the preliminary synthetic data to obtain synthetic data.
[0131] Through the trained ground penetrating radar deep convolutional generative adversarial network, the weights of the generator trained in it are used to optimize the preliminary synthetic data, thereby obtaining more realistic synthetic data.
[0132] The above-mentioned optimized detection model is trained by using the obtained synthetic data and real data, so as to obtain a trained optimized detection model.
[0133] In this application, an embodiment of the above-mentioned single-stage lightweight subway lining cavity identification method is also provided. Specifically, data collection is collected using antenna frequencies of 800MHz and 500MHz, and the collection area is a section of subway railway that was once constructed using the mining method. Fig.10 and Fig.11 The data collection process and tunnel structure are shown. The radar antenna has a track spacing of 0.02m, a sampling frequency of 12000MHz, a time window of 63ns, and a collection line of left and right arch waists. When collecting data, the ground penetrating radar antenna is close to the tunnel lining surface and the radar antenna is pushed to move. The radar system will detect underground objects by sending wireless pulse signals and recording the received echo signals. After obtaining the tunnel cavity disease data collected by the ground penetrating radar, the above tunnel cavity disease data is preprocessed to obtain preprocessed data for tunnel inspection.
[0134] Afterwards, in order to evaluate the impact of four different optimization methods on the YOLOv7 network model in the single-stage lightweight subway lining cavity recognition method, and to understand the optimal effect achieved by the optimized detection model, the embodiment of the present application conducted an ablation experiment to process the pre-processed data obtained above, including five groups of experiments, each of which added different modules for testing and compared with the original YOLOv7 model. The performance of the model is measured using indicators such as F1 score (F1Score, F1 score), number of parameters, floating point operations (Floating Point Operations, referred to as FLOPs), FPS (Frames Per Second, FPS means the number of image frames processed per second, used to measure the reasoning speed of the model, a higher FPS value means that the model can reason faster) and mean average precision (mean average precision, mAP) under a threshold of 0.5. Among them, the YOLOv7 model with the ACmix module is called YOLOv7+ACmix, the model with a pure convolutional module backbone network is called YOLOv7+CXNB, the model with the PELAN module is called YOLOv7+PELAN, the model with the WIoU loss function is called YOLOv7+WIoU, and the optimized detection model is called CAPW-YOLO.
[0135] The ablation experiment was conducted using a computing system equipped with an NVIDIA GeForce RTX A6000 GPU with 48GB VRAM; the operating system used was Windows 10, and Python 3.8 was used as the software development framework. The training parameters used for the model in the experiment are recorded in Tables 2 and 3, respectively.
[0136] parameter value parameter value Learning Rate 0.001 Batch size 32 Weight decay 0.0005 Image size 800×320 momentum 0.937 Iteration cycle 200
[0137] Table 2. Experimental environment
[0138]
[0139] Table 3. Ablation experiment
[0140] As shown in Table 3, after adding the CNXB module to YOLOv7, the average accuracy at a threshold of 0.5 increased by 4.8%, and the F1 score increased by 3%; at the same time, the floating-point operation volume decreased significantly by 60%, which shows that the CNXB module can not only improve the accuracy of the optimized detection model, but also greatly reduce the floating-point operation volume of the optimized detection model, making the optimized detection model more lightweight. After adding the ACmix module, the average accuracy at a threshold of 0.5 improved by 1.4%, and the F1 score increased by 1%; however, adding the ACmix module also led to an increase in parameters by 11% and an increase in floating-point operations by 18.9%. Although adding the ACmix module has a positive impact on the accuracy of the model, its implementation is associated with increased computational requirements and reduced efficiency. Adding the PELAN module to the YOLOv7 model shows a significant improvement in model performance, with an average accuracy at a threshold of 0.5 increased by 2.7%, an F1 score increased by 3%, and a significant reduction in the complexity of the model; specifically, the PELAN module successfully reduced parameters by 23.9% and floating-point operations by 28.4%. These results show that the PELAN module can effectively improve the accuracy of the model while reducing the computational cost. It can also be seen from Table 3 above that the optimized detection model can detect small-scale defects with unclear features that YOLOv7 fails to detect, and also shows significant improvement in detecting defects with prominent void features.
[0141] It can be seen from the above that compared with the prior art, the use of a single-stage method for detecting tunnel cavity disease data still has shortcomings, resulting in inaccurate detection results. The single-stage lightweight subway lining cavity recognition method proposed in this application uses the optimized detection model obtained by optimizing the feature extraction and fusion process and loss function of the YOLOv7 (one of the YOLO detection models, referred to as YOLOv7) detection model to process the tunnel cavity disease data, thereby enhancing the process of extracting local features, global features and fusion features of the optimized detection model, and at the same time reducing the computational cost without affecting the performance of the model. The accuracy of the model is improved, and the robustness of the model is improved by optimizing the loss function; and in the present application, the training data used in training the optimized detection model of the model is synthetic data obtained by optimizing the deep convolution generative adversarial network of the ground penetrating radar, so that the generated synthetic data is more formal, and because the optimized detection model is trained by a combination of synthetic data and real data, the number of training data sets is greatly increased, thereby improving the training effect of the optimized detection model, and a more accurate trained optimized detection model can be obtained, so that the obtained optimized detection model can obtain effective and accurate detection results in a short time.
[0142] Exemplary Devices
[0143] like Fig.12 As shown in , corresponding to the above-mentioned single-stage lightweight subway lining cavity identification method, the embodiment of the present invention further provides a single-stage lightweight subway lining cavity identification system, and the above-mentioned single-stage lightweight subway lining cavity identification system includes:
[0144] The data preprocessing module 51 is used to obtain tunnel cavity disease data, and perform tunnel cavity preprocessing on the tunnel cavity disease data to obtain preprocessed data, wherein the tunnel disease data at least includes tunnel cavity location information and tunnel cavity size information;
[0145] The data processing module 52 is used to input the preprocessed data into the trained optimization detection model to obtain and output a picture with a hole detection result, wherein the optimization detection model is trained by real data and synthetic data obtained by optimizing the ground penetrating radar deep convolution adversarial network.
[0146] It should be noted that the specific structure and implementation of the above-mentioned single-stage lightweight subway lining void identification system and its various modules or units can refer to the corresponding description in the above-mentioned method embodiment, and will not be repeated here.
[0147] It should be noted that the division method of each module of the above-mentioned single-stage lightweight subway lining cavity identification system is not unique and is not specifically limited here.
[0148] Based on the above embodiments, the present invention further provides an intelligent terminal, whose principle block diagram can be shown as follows: Fig.13 As shown. The above-mentioned intelligent terminal includes a processor, a memory, a network interface and a display screen connected through a system bus. Among them, the processor of the intelligent terminal is used to provide computing and control capabilities. The memory of the intelligent terminal includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a single-stage lightweight subway lining void identification program. The internal memory provides an environment for the operation of the operating system and the single-stage lightweight subway lining void identification program in the non-volatile storage medium. The network interface of the intelligent terminal is used to communicate with an external terminal through a network connection. When the single-stage lightweight subway lining void identification program is executed by the processor, the steps of any one of the above-mentioned single-stage lightweight subway lining void identification methods are implemented. The display screen of the intelligent terminal can be a liquid crystal display screen or an electronic ink display screen.
[0149] Those skilled in the art will understand that Fig.13 The principle block diagram shown in the figure is only a block diagram of a partial structure related to the scheme of the present invention, and does not constitute a limitation on the smart terminal to which the scheme of the present invention is applied. The specific smart terminal may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.
[0150] In one embodiment, a smart terminal is provided, which includes a memory, a processor, and a single-stage lightweight subway lining void identification program stored in the memory and executable on the processor. When the single-stage lightweight subway lining void identification program is executed by the processor, the steps of any one of the single-stage lightweight subway lining void identification methods provided in the embodiments of the present invention are implemented.
[0151] An embodiment of the present invention further provides a computer-readable storage medium, on which a single-stage lightweight subway lining void identification program is stored. When the single-stage lightweight subway lining void identification program is executed by a processor, the steps of any single-stage lightweight subway lining void identification method provided in an embodiment of the present invention are implemented.
[0152] It should be understood that the serial numbers of the steps in the above embodiments do not imply a sequence of execution, and the execution sequence of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.
[0153] Those skilled in the art can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the above-mentioned device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of the present invention. The specific working process of the units and modules in the above-mentioned device can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.
[0154] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.
[0155] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.
[0156] In the embodiments provided by the present invention, it should be understood that the disclosed system / terminal device and method can be implemented in other ways. For example, the system / terminal device embodiments described above are only illustrative, for example, the division of the above modules or units is only a logical function division, and in actual implementation, other division methods can be used, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.
[0157] If the above-mentioned integrated module / unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The above-mentioned computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Among them, the above-mentioned computer program includes computer program code, and the above-mentioned computer program code can be in source code form, object code form, executable file or some intermediate form. The above-mentioned computer-readable medium may include: any entity or device capable of carrying the above-mentioned computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium. It should be noted that the content contained in the above-mentioned computer-readable storage medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction.
[0158] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, it should be understood by those skilled in the art that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features therein may be replaced by equivalents. However, these modifications or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention.
Claims
1. A single-stage lightweight subway lining cavity identification method, characterized in that: The single-stage lightweight subway lining cavity identification method comprises: Acquire tunnel cavity disease data, and perform tunnel cavity preprocessing on the tunnel cavity disease data to obtain preprocessed data, wherein the tunnel disease data at least includes tunnel cavity location information and tunnel cavity size information; Inputting the preprocessed data into a trained optimization detection model to obtain and output a picture with a hole detection result, wherein the optimization detection model is trained by using real data and synthetic data optimized by a ground penetrating radar deep convolution generative adversarial network; The step of generating the optimized detection model comprises: Based on the YOLOv7 detection model, the optimized detection model is obtained by optimizing the feature extraction and fusion process and optimizing the loss function; The optimization of the feature extraction and fusion process specifically includes: Optimizing the feature extraction and fusion process through a pure convolution module, a fusion module of self-attention and convolution, and an optimized efficient aggregation network; The pure convolution module contains pure convolution blocks. The four 3×3 convolutions in the ELAN module are replaced by the ConvNeXt Block module. A CNXB module and a CBS module are used to replace the first ELAN module in the backbone network. The pure convolution block extracts and normalizes features through a 7×7 deep convolution layer and a normalization layer. The channel size is quadrupled and nonlinearity is introduced through 1×1 convolution and GE LU activation function. The channel size is restored to its original size through 1×1 convolution, layer scaling and Drop Path layer. Finally, the input and output are added together. The fusion module of self-attention and convolution is divided into two stages, stage I and stage II; in stage I, the input features are projected by convolution and reshaped into N blocks respectively to obtain an intermediate feature set containing feature maps; in stage II, the structure is divided into two paths, the self-attention path and the convolution path; in the self-attention path, the intermediate feature set is clustered into N groups, each containing 3 feature maps, which are used as query, key and value respectively, and are processed according to the traditional multi-head self-attention model; in the convolution path, a lightweight fully connected layer is used to generate N groups, each with K 2 feature maps; the feature maps generated in the two stages are processed by translation and aggregation, the input features are processed in a convolutional manner and information is collected from the local receptive field. Finally, the outputs of the self-attention and convolution paths are summed; The optimized efficient aggregation network uses a combination of parallel convolution and 1x1CBS modules to replace the CBS module in the original efficient aggregation network based on the efficient aggregation network; based on YOLOv7, the first ELAN module is replaced by a pure convolution module and a CBS module, and all other ELAN modules are replaced by the optimized efficient aggregation network.
2. The single-stage lightweight subway lining cavity identification method according to claim 1 is characterized in that: The step of obtaining tunnel cavity disease data and performing tunnel cavity preprocessing on the tunnel cavity disease data to obtain preprocessed data specifically includes: Acquiring the tunnel cavity disease data by radar; The tunnel cavity disease data is subjected to tunnel cavity preprocessing to obtain the preprocessed data, wherein the preprocessing includes at least one of direct current offset elimination, gain adjustment, time zero point correction, bandpass filtering and sliding average operation.
3. The single-stage lightweight subway lining cavity identification method according to claim 1 is characterized in that: The optimization of the loss function specifically includes: The loss function is optimized by a weighted intersection-over-union function.
4. The single-stage lightweight subway lining cavity identification method according to claim 1 is characterized in that: The optimized detection model is trained by using real data and synthetic data optimized by a ground penetrating radar deep convolution generative adversarial network, specifically including: Preliminary synthetic data were generated based on the finite-difference time-domain method; The preliminary synthetic data is optimized according to the trained ground penetrating radar deep convolutional generative adversarial network to obtain the synthetic data.
5. The single-stage lightweight subway lining cavity identification method according to claim 4 is characterized in that: The ground penetrating radar deep convolutional generative adversarial network includes a generator network and a discriminator network. The ground penetrating radar deep convolutional generative adversarial network is trained by the generator network and the discriminator network to obtain the trained ground penetrating radar deep convolutional generative adversarial network.
6. A single-stage lightweight subway lining cavity identification system, characterized in that: The single-stage lightweight subway lining void identification system includes: A data preprocessing module, used for acquiring tunnel cavity disease data, and performing tunnel cavity preprocessing on the tunnel cavity disease data to obtain preprocessed data, wherein the tunnel disease data at least includes tunnel cavity location information and tunnel cavity size information; A data processing module, used for inputting the pre-processed data into a trained optimization detection model to obtain and output a picture with a hole detection result, wherein the optimization detection model is trained by real data and synthetic data optimized by a ground penetrating radar deep convolution generative adversarial network; The step of generating the optimized detection model comprises: Based on the YOLOv7 detection model, the optimized detection model is obtained by optimizing the feature extraction and fusion process and optimizing the loss function; The optimization of the feature extraction and fusion process specifically includes: Optimizing the feature extraction and fusion process through a pure convolution module, a fusion module of self-attention and convolution, and an optimized efficient aggregation network; The pure convolution module contains pure convolution blocks. The four 3×3 convolutions in the ELAN module are replaced by the ConvNeXt Block module. A CNXB module and a CBS module are used to replace the first ELAN module in the backbone network. The pure convolution block extracts and normalizes features through a 7×7 deep convolution layer and a normalization layer. The channel size is quadrupled and nonlinearity is introduced through 1×1 convolution and GE LU activation function. The channel size is restored to its original size through 1×1 convolution, layer scaling and Drop Path layer. Finally, the input and output are added together. The fusion module of self-attention and convolution is divided into two stages, stage I and stage II; in stage I, the input features are projected by convolution and reshaped into N blocks respectively to obtain an intermediate feature set containing feature maps; in stage II, the structure is divided into two paths, the self-attention path and the convolution path; in the self-attention path, the intermediate feature set is clustered into N groups, each containing 3 feature maps, which are used as query, key and value respectively, and are processed according to the traditional multi-head self-attention model; in the convolution path, a lightweight fully connected layer is used to generate N groups, each with K 2 feature maps; the feature maps generated in the two stages are processed by translation and aggregation, the input features are processed in a convolutional manner and information is collected from the local receptive field. Finally, the outputs of the self-attention and convolution paths are summed; The optimized efficient aggregation network uses a combination of parallel convolution and 1x1CBS modules to replace the CBS module in the original efficient aggregation network based on the efficient aggregation network; based on YOLOv7, the first ELAN module is replaced by a pure convolution module and a CBS module, and all other ELAN modules are replaced by the optimized efficient aggregation network.
7. An intelligent terminal, characterized in that: The intelligent terminal includes a memory, a processor, and a single-stage lightweight subway lining void identification program stored in the memory and executable on the processor. When the single-stage lightweight subway lining void identification program is executed by the processor, the steps of the single-stage lightweight subway lining void identification method according to any one of claims 1 to 5 are implemented.
8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a single-stage lightweight subway lining void identification program, which, when executed by a processor, implements the steps of the single-stage lightweight subway lining void identification method as described in any one of claims 1 to 5.
