An Automatic Identification Method and System for Tunnel Water Inrush Based on a Target Detection Model

Abstract

This invention provides an automatic identification method and system for tunnel inrush water based on an inrush water target detection model, belonging to the field of tunnel inrush water disaster monitoring and prevention technology. The invention extracts features from video images within the tunnel; updates the features by mixing spatial attention weights and channel attention weights; fuses the updated features; and decodes the fused features to obtain the target bounding box and category result. By employing a content-aware attention mechanism and coupling contextual features, this invention improves the generalization ability of the monitoring model, ensuring accurate identification of tunnel inrush water under various tunnel conditions. Furthermore, it achieves comprehensive, high-precision detection and statistical analysis of tunnel inrush water, determining whether to issue early warnings within the tunnel, thereby predicting the evolution of the inrush water and providing assurance for safe tunnel construction.

Description

Technical Field

[0001] This invention relates to the field of tunnel water inrush disaster monitoring and prevention technology, and in particular to an automatic identification method and system for tunnel water inrush based on a water inrush target detection model. Background Technology

[0002] During the construction of deep-buried tunnels using the drill-and-blast method, hazards such as sudden water inrushes often arise due to traversing high mountain canyons, water-rich areas, fault fracture zones, and karst caves. Given the complexity of the geological conditions in the tunnel construction area and the uncertainties inherent in the construction process, accurate monitoring of sudden water inrushes is crucial for ensuring safe tunnel construction and preventing casualties. Monitoring is an indispensable step in handling tunnel sudden water inrush hazards.

[0003] The patent "Method, Device and System for Unattended Monitoring and Early Warning of Tunnel Water Inrush Using Machine Vision", patent number "CN202310786056.3", uses the method of calculating optical flow vectors to calculate tunnel water inrush. This method is easily affected by light and dust. When there is a lot of dust in the tunnel, the accuracy of the method is greatly reduced, and the staff cannot accurately find the water inrush point from the video.

[0004] Therefore, a method for accurately identifying sudden water inrushes in tunnels is needed. Summary of the Invention

[0005] In view of this, the present invention provides an automatic identification method and system for tunnel inrush water based on an inrush water target detection model. By constructing an inrush water target detection model, the automatic identification of inrush water is ensured under various tunnel conditions, and the comprehensive high-precision detection and statistical analysis of tunnel inrush water is realized.

[0006] Therefore, the present invention provides the following technical solution:

[0007] An automatic identification method for tunnel water inrush based on a water inrush target detection model includes:

[0008] Construct a target detection model for sudden water inrush;

[0009] The sudden water inrush target detection model includes:

[0010] Extract features from video images inside the tunnel;

[0011] The spatial attention weights and channel attention weights of the mixed features are used to update the features;

[0012] Integrate the updated features;

[0013] The fused features are decoded to obtain the target bounding box and category results.

[0014] Furthermore, the extraction of video image features within the tunnel includes:

[0015] Four layers of features were obtained by feature extraction using a ResNet50 network.

[0016] The encoder obtains a new feature from the fourth layer of the four-layer features.

[0017] Furthermore, the spatial attention weights and channel attention weights of the hybrid features update the features, including:

[0018] The new features of the fourth layer and the channel attention weights of the first, second and third layer features in the four layers are obtained by convolution, global average pooling and channel shuffling.

[0019] The new features of the fourth layer features and the spatial attention weights of the first, second and third layer features in the four layers are obtained by global average pooling, global max pooling and channel shuffling.

[0020] The channel attention weights and the spatial attention weights are combined;

[0021] Update features using a content-aware attention mechanism.

[0022] Furthermore, the updated features are fused using a multi-layer feature fusion module:

[0023]

[0024] f l =f l-1 +f l+1 l = 7

[0025] f k =f k-4 +DS(f k-1 ), k = 8, 9

[0026] F = PConv(f) 10 )+PConv(f9)+PConv(DS(f8))+PConv(DS(f7)

[0027] Where PConv represents a partial convolution operation, f i The features are represented by DS, which represents the downsampling operation, and BL represents the bilinear interpolation upsampling operation.

[0028] Furthermore, the decoding of the fused features to obtain the target bounding box and category results includes:

[0029] The fused features F are used to obtain the bounding box and category results through the transformer decoder module and the feedforward neural network.

[0030] Furthermore, the category results include:

[0031] Sudden water inrush point and sudden water inrush.

[0032] Furthermore, before constructing the sudden water inrush target detection model, the method further includes: constructing a training dataset for the sudden water inrush target detection model.

[0033] Furthermore, the training dataset for constructing the sudden water inrush target detection model includes:

[0034] Acquire video frames of the sudden water inrush and perform preprocessing;

[0035] The preprocessed video frames were labeled and segmented to serve as the training dataset for the sudden water inrush target detection model.

[0036] The preprocessing includes:

[0037] Extract video frames of the sudden water inrush from the surveillance video inside the tunnel;

[0038] Wavelet denoising was used to decompose the video frames of the sudden water surge into subbands of different frequencies;

[0039] Thresholding is applied to subbands of different frequencies to remove noise from the video frames of sudden water inrush.

[0040] An automatic identification system for tunnel water inrush based on a water inrush target detection model includes:

[0041] The main module is used to extract features from video images inside the tunnel;

[0042] The content-aware module uses spatial attention weights and channel attention weights to mix features and update the features.

[0043] A multi-layer feature fusion module fuses updated features.

[0044] The recognition module decodes the fused features to obtain the target bounding box and category results.

[0045] Advantages and positive effects of the present invention:

[0046] This invention utilizes industrial cameras to monitor tunnel video in real time and employs a sudden water inrush detection target model to detect targets in the monitored video images, automatically identifying sudden water inrush points and water inrushes. The invention detects locations where sudden water inrushes may occur, such as the tunnel face and arch waist, and includes supplementary lighting to enhance video brightness. For the sudden water inrush target detection model in this invention, frames from actual sudden water inrush videos are acquired to construct a training set for a realistic sudden water inrush target detection model. A content-aware attention mechanism is used to acquire specific important spatial information in each channel, ensuring sufficient mixing of spatial and channel attention features to guarantee information interaction. This addresses the problem that sudden water inrush points in tunnels are small targets in images, occupying a small proportion of the total pixels, and the foreground and background colors are nearly identical, making target capture difficult. This improves the generalization ability of the monitoring model, ensuring accurate identification of sudden water inrushes under various tunnel conditions. Furthermore, it achieves comprehensive, high-precision detection and statistical analysis of sudden water inrushes in tunnels, determining whether to issue early warnings within the tunnel, thereby predicting the evolution of sudden water inrushes and providing assurance for safe tunnel construction. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a flowchart of the method in an embodiment of the present invention;

[0049] Figure 2 This is a diagram showing the equipment layout in an embodiment of the present invention;

[0050] Figure 3 This is a framework diagram of the front-end display system in an embodiment of the present invention;

[0051] Figure 4 This is a framework diagram of the sudden water inrush target detection model in an embodiment of the present invention;

[0052] Figure 5 This is a framework diagram of the CA_A content-aware attention module in an embodiment of the present invention;

[0053] Figure 6 This is a diagram showing the detection results in an embodiment of the present invention. Detailed Implementation

[0054] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0055] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0056] This invention provides an automatic identification method for tunnel water inrush based on a water inrush target detection model. It utilizes video frames of tunnel water inrush acquired within the tunnel to construct a water inrush target detection model to obtain the location and quantity of the water inrush; it provides early warning alerts within the tunnel, achieving real-time monitoring of tunnel water inrush; specifically, it combines... Figure 4 The method steps of the present invention include:

[0057] S1. Feature extraction is performed using a ResNet50 network;

[0058] S2. Feed the last layer feature F5 into the Transformer encoder module to obtain the new feature S5;

[0059] Q, K, V = Flatten (F5)

[0060] S5=Reshape(Attention(Q,K,V))

[0061] Where Q (Query) is the query vector, representing the information that the current element needs to obtain from other elements.

[0062] K (Key): The key vector, used to match the query vector to determine which elements are relevant to the current element.

[0063] V(Value): A numerical vector containing the actual content of the selected element, which is weighted and summed according to the degree of matching with the query vector.

[0064] Flatten represents the transformation of multidimensional features into one dimension, while the Reshape operation represents the reshaping of the dimensions of features;

[0065] S3, the new feature S5 of the updated 4th layer, and the features of the first three layers, the specific structure is as follows: Figure 5 As shown;

[0066] S31. Obtain an exclusive spatial importance map for each single channel of the input feature in a coarse-to-fine manner, and fully mix channel attention weights and spatial attention weights:

[0067]

[0068] W out =W s +W c +CS(C 1*1 (Y))

[0069] Where CS represents channel shuffling; GAP represents global average pooling; GMP represents global max pooling; max(0,x) represents the ReLU activation function; C k*k W represents a convolution with a kernel size of k*k; c Represents spatial attention weights, W s This represents the channel attention weight.

[0070] S32. Based on the spatial attention output value and the channel attention output value, the four-layer features will be processed by the content-aware module to obtain new four-layer features:

[0071]

[0072] Here, CA_A represents content-aware attention.

[0073] S4. The four layers of features after passing through the content-aware module are fused to promote semantic communication between features at different levels, enrich the semantic information at each level, and enable accurate identification of small water inrush points in the tunnel. The specific feature fusion formula is as follows:

[0074]

[0075] f l =f l-1 +f l+1 l = 7

[0076] f k =f k-4 +DS(f k-1 ), k = 8, 9

[0077] The lower two layers are then downsampled to match the dimensionality of the higher-level features. After partial convolutional PConv operations, the features are fused to obtain the final output feature F.

[0078] F = PConv(f) 10 )+PConv(f9)+PConv(DS(f8))+PConv(DS(f7)

[0079] Where PConv represents a partial convolution operation, f i The features are represented by DS, which represents the downsampling operation, and BL represents the bilinear interpolation upsampling operation.

[0080] S5. The feature F after feature fusion is fed into the transformer decoder module, and then passed through the feedforward neural network to finally obtain the target box and category results.

[0081] This invention also provides an automatic identification system for tunnel water inrush based on a water inrush target detection model, combined with... Figure 4 Further explanation:

[0082] The main module is used to extract features from video images inside the tunnel;

[0083] The content-aware module uses spatial attention weights and channel attention weights to mix features and update the features.

[0084] A multi-layer feature fusion module fuses updated features.

[0085] The recognition module decodes the fused features to obtain the target bounding box and category results.

[0086] Combination Figure 1 The method and system of the present invention will be further illustrated with specific application examples:

[0087] Equipment deployment, such as Figure 2 As shown.

[0088] In this embodiment, preferably, an industrial camera with a field of view of more than 10m and a pixel count of no less than 12 million is mounted on the tunnel arch to capture clear video of the sudden water inrush in the tunnel in real time; and a supplementary light is set above the industrial camera to improve the light intensity.

[0089] Collect training datasets for the sudden water inrush target detection model;

[0090] The video is transmitted to the video acquisition module of the front-end display system.

[0091] The video from the video acquisition module of the front-end display system is transmitted to the local server;

[0092] Video frames with obvious water inrushes are extracted, and wavelet denoising is used to decompose the image into sub-bands of different frequencies. Thresholding is applied to the sub-bands, and appropriate wavelet bases and decomposition levels are selected to remove noise, thereby achieving data enhancement of the water inrush video frames.

[0093] The enhanced video frames of sudden water inrush were labeled using the Labelme annotation tool. The labels were divided into two categories: one for sudden water inrush points and the other for sudden water inrush. The training set and test set of the dataset were divided in an 8:1 ratio.

[0094] The sudden water inrush target detection model is pre-trained based on the sudden water inrush training set, and then the pre-trained sudden water inrush target detection model is used to detect sudden water inrush.

[0095] The training results and target recognition results are output to the front-end display page for display and warning; the number of sudden water inrush points is counted to determine the current water inrush situation in order to provide an early warning on whether tunneling can continue and to ensure the safety of construction personnel.

[0096] In this embodiment, the output detection result is as follows: Figure 6 As shown.

[0097] Combination Figure 3 As shown, the system in this embodiment operates and displays based on a computing device, including:

[0098] The system of this invention uses a system registration module to register and log in as a basic user and an administrator;

[0099] The video acquisition and storage module in the computing device acquires and displays the video streams of the tunnel face and sidewalls in real time; and saves the video frames within a certain period of time to the local server.

[0100] In this invention, the sudden water inrush target detection model is implemented based on a detection module, which uploads real-time video and detects and counts the number of sudden water inrush targets in the real-time video; and through a settings module, it manages the storage path, storage format, and storage quality of the target detection results of the detection module.

[0101] This invention obtains real-time monitoring videos of locations where sudden water inrush may occur, such as the tunnel face and arch, and creates a training set from the video frames showing obvious water inrushes. Then, it detects sudden water inrushes through target detection. The invention uses a content-aware attention mechanism coupled with contextual features to ensure accurate identification of tunnel sudden water inrushes under various tunnel conditions. It achieves comprehensive, high-precision detection and statistical analysis of tunnel sudden water inrushes and determines whether to issue early warnings within the tunnel, thereby predicting the evolution of the sudden water inrush and providing a guarantee for safe tunnel construction.

[0102] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An automatic identification method for tunnel inrush water based on an inrush water target detection model, characterized in that, include: Construct a target detection model for sudden water inrush; The sudden water inrush target detection model includes: Extracting features from video images inside the tunnel: Four layers of features were obtained by feature extraction using a ResNet50 network. The encoder obtains a new feature from the fourth layer of the four-layer features; The spatial attention weights and channel attention weights of the hybrid features are used to update features, including: The new features of the fourth layer and the channel attention weights of the first, second and third layer features in the four layers are obtained by convolution, global average pooling and channel shuffling. The new features of the fourth layer features and the spatial attention weights of the first, second and third layer features in the four layers are obtained by global average pooling, global max pooling and channel shuffling. For each single channel of the input feature, an exclusive spatial importance map is obtained in a coarse-to-fine manner, fully blending channel attention weights and spatial attention weights: in, This indicates a channel shuffling operation; Indicates global average pooling; Indicates global max pooling; express Activation function; Represents a convolution with a kernel size of k*k; Represents spatial attention weights, Indicates channel attention weights; Update features through content-aware attention mechanisms; The updated features are fused using a multi-layer feature fusion module: in, This indicates a partial convolution operation. The features are represented by DS, which represents the downsampling operation, and BL, which represents the bilinear interpolation upsampling operation. The fused features are decoded to obtain the target bounding box and category results.

2. The method for automatic identification of tunnel inrush water based on an inrush water target detection model according to claim 1, characterized in that, The process of decoding the fused features to obtain the target bounding box and category results includes: Features after fusion The target bounding box and category results are obtained through the transformer decoder module and the feedforward neural network.

3. The method for automatic identification of tunnel inrush water based on an inrush water target detection model according to claim 2, characterized in that, The category results include: water inrush points and water inrush.

4. The method for automatic identification of tunnel inrush water based on an inrush water target detection model according to claim 1, characterized in that, Before constructing the target detection model for sudden water inrush, the following steps are also taken: constructing a training dataset for the target detection model for sudden water inrush.

5. The method for automatic identification of tunnel inrush water based on an inrush water target detection model according to claim 4, characterized in that, The training dataset for constructing the sudden water inrush target detection model includes: Acquire video frames of the sudden water inrush and perform preprocessing; The preprocessed video frames are labeled and segmented to serve as the training dataset for the sudden water inrush target detection model; the preprocessing includes: Extract video frames of the sudden water inrush from the surveillance video inside the tunnel; Wavelet denoising was used to decompose the video frames of the sudden water surge into subbands of different frequencies; Thresholding is applied to subbands of different frequencies to remove noise from the video frames of sudden water inrush.

6. An automatic tunnel water inrush detection system based on the water inrush target detection model described in any one of claims 1-5, characterized in that, include: The main module is used to extract features from video images inside the tunnel; The content-aware module updates features using spatial attention weights and channel attention weights that combine features; A multi-layer feature fusion module fuses updated features. The recognition module decodes the fused features to obtain the target bounding box and category results.

Images